Commit | Line | Data |
---|---|---|
c9de560d AT |
1 | /* |
2 | * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com | |
3 | * Written by Alex Tomas <alex@clusterfs.com> | |
4 | * | |
5 | * This program is free software; you can redistribute it and/or modify | |
6 | * it under the terms of the GNU General Public License version 2 as | |
7 | * published by the Free Software Foundation. | |
8 | * | |
9 | * This program is distributed in the hope that it will be useful, | |
10 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
11 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
12 | * GNU General Public License for more details. | |
13 | * | |
14 | * You should have received a copy of the GNU General Public Licens | |
15 | * along with this program; if not, write to the Free Software | |
16 | * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111- | |
17 | */ | |
18 | ||
19 | ||
20 | /* | |
21 | * mballoc.c contains the multiblocks allocation routines | |
22 | */ | |
23 | ||
18aadd47 | 24 | #include "ext4_jbd2.h" |
8f6e39a7 | 25 | #include "mballoc.h" |
28623c2f | 26 | #include <linux/log2.h> |
a0b30c12 | 27 | #include <linux/module.h> |
5a0e3ad6 | 28 | #include <linux/slab.h> |
9bffad1e TT |
29 | #include <trace/events/ext4.h> |
30 | ||
a0b30c12 TT |
31 | #ifdef CONFIG_EXT4_DEBUG |
32 | ushort ext4_mballoc_debug __read_mostly; | |
33 | ||
34 | module_param_named(mballoc_debug, ext4_mballoc_debug, ushort, 0644); | |
35 | MODULE_PARM_DESC(mballoc_debug, "Debugging level for ext4's mballoc"); | |
36 | #endif | |
37 | ||
c9de560d AT |
38 | /* |
39 | * MUSTDO: | |
40 | * - test ext4_ext_search_left() and ext4_ext_search_right() | |
41 | * - search for metadata in few groups | |
42 | * | |
43 | * TODO v4: | |
44 | * - normalization should take into account whether file is still open | |
45 | * - discard preallocations if no free space left (policy?) | |
46 | * - don't normalize tails | |
47 | * - quota | |
48 | * - reservation for superuser | |
49 | * | |
50 | * TODO v3: | |
51 | * - bitmap read-ahead (proposed by Oleg Drokin aka green) | |
52 | * - track min/max extents in each group for better group selection | |
53 | * - mb_mark_used() may allocate chunk right after splitting buddy | |
54 | * - tree of groups sorted by number of free blocks | |
55 | * - error handling | |
56 | */ | |
57 | ||
58 | /* | |
59 | * The allocation request involve request for multiple number of blocks | |
60 | * near to the goal(block) value specified. | |
61 | * | |
b713a5ec TT |
62 | * During initialization phase of the allocator we decide to use the |
63 | * group preallocation or inode preallocation depending on the size of | |
64 | * the file. The size of the file could be the resulting file size we | |
65 | * would have after allocation, or the current file size, which ever | |
66 | * is larger. If the size is less than sbi->s_mb_stream_request we | |
67 | * select to use the group preallocation. The default value of | |
68 | * s_mb_stream_request is 16 blocks. This can also be tuned via | |
69 | * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in | |
70 | * terms of number of blocks. | |
c9de560d AT |
71 | * |
72 | * The main motivation for having small file use group preallocation is to | |
b713a5ec | 73 | * ensure that we have small files closer together on the disk. |
c9de560d | 74 | * |
b713a5ec TT |
75 | * First stage the allocator looks at the inode prealloc list, |
76 | * ext4_inode_info->i_prealloc_list, which contains list of prealloc | |
77 | * spaces for this particular inode. The inode prealloc space is | |
78 | * represented as: | |
c9de560d AT |
79 | * |
80 | * pa_lstart -> the logical start block for this prealloc space | |
81 | * pa_pstart -> the physical start block for this prealloc space | |
53accfa9 TT |
82 | * pa_len -> length for this prealloc space (in clusters) |
83 | * pa_free -> free space available in this prealloc space (in clusters) | |
c9de560d AT |
84 | * |
85 | * The inode preallocation space is used looking at the _logical_ start | |
86 | * block. If only the logical file block falls within the range of prealloc | |
caaf7a29 TM |
87 | * space we will consume the particular prealloc space. This makes sure that |
88 | * we have contiguous physical blocks representing the file blocks | |
c9de560d AT |
89 | * |
90 | * The important thing to be noted in case of inode prealloc space is that | |
91 | * we don't modify the values associated to inode prealloc space except | |
92 | * pa_free. | |
93 | * | |
94 | * If we are not able to find blocks in the inode prealloc space and if we | |
95 | * have the group allocation flag set then we look at the locality group | |
caaf7a29 | 96 | * prealloc space. These are per CPU prealloc list represented as |
c9de560d AT |
97 | * |
98 | * ext4_sb_info.s_locality_groups[smp_processor_id()] | |
99 | * | |
100 | * The reason for having a per cpu locality group is to reduce the contention | |
101 | * between CPUs. It is possible to get scheduled at this point. | |
102 | * | |
103 | * The locality group prealloc space is used looking at whether we have | |
25985edc | 104 | * enough free space (pa_free) within the prealloc space. |
c9de560d AT |
105 | * |
106 | * If we can't allocate blocks via inode prealloc or/and locality group | |
107 | * prealloc then we look at the buddy cache. The buddy cache is represented | |
108 | * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets | |
109 | * mapped to the buddy and bitmap information regarding different | |
110 | * groups. The buddy information is attached to buddy cache inode so that | |
111 | * we can access them through the page cache. The information regarding | |
112 | * each group is loaded via ext4_mb_load_buddy. The information involve | |
113 | * block bitmap and buddy information. The information are stored in the | |
114 | * inode as: | |
115 | * | |
116 | * { page } | |
c3a326a6 | 117 | * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... |
c9de560d AT |
118 | * |
119 | * | |
120 | * one block each for bitmap and buddy information. So for each group we | |
121 | * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE / | |
122 | * blocksize) blocks. So it can have information regarding groups_per_page | |
123 | * which is blocks_per_page/2 | |
124 | * | |
125 | * The buddy cache inode is not stored on disk. The inode is thrown | |
126 | * away when the filesystem is unmounted. | |
127 | * | |
128 | * We look for count number of blocks in the buddy cache. If we were able | |
129 | * to locate that many free blocks we return with additional information | |
130 | * regarding rest of the contiguous physical block available | |
131 | * | |
132 | * Before allocating blocks via buddy cache we normalize the request | |
133 | * blocks. This ensure we ask for more blocks that we needed. The extra | |
134 | * blocks that we get after allocation is added to the respective prealloc | |
135 | * list. In case of inode preallocation we follow a list of heuristics | |
136 | * based on file size. This can be found in ext4_mb_normalize_request. If | |
137 | * we are doing a group prealloc we try to normalize the request to | |
27baebb8 TT |
138 | * sbi->s_mb_group_prealloc. The default value of s_mb_group_prealloc is |
139 | * dependent on the cluster size; for non-bigalloc file systems, it is | |
c9de560d | 140 | * 512 blocks. This can be tuned via |
d7a1fee1 | 141 | * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in |
c9de560d AT |
142 | * terms of number of blocks. If we have mounted the file system with -O |
143 | * stripe=<value> option the group prealloc request is normalized to the | |
d7a1fee1 DE |
144 | * the smallest multiple of the stripe value (sbi->s_stripe) which is |
145 | * greater than the default mb_group_prealloc. | |
c9de560d | 146 | * |
d7a1fee1 | 147 | * The regular allocator (using the buddy cache) supports a few tunables. |
c9de560d | 148 | * |
b713a5ec TT |
149 | * /sys/fs/ext4/<partition>/mb_min_to_scan |
150 | * /sys/fs/ext4/<partition>/mb_max_to_scan | |
151 | * /sys/fs/ext4/<partition>/mb_order2_req | |
c9de560d | 152 | * |
b713a5ec | 153 | * The regular allocator uses buddy scan only if the request len is power of |
c9de560d AT |
154 | * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The |
155 | * value of s_mb_order2_reqs can be tuned via | |
b713a5ec | 156 | * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to |
af901ca1 | 157 | * stripe size (sbi->s_stripe), we try to search for contiguous block in |
b713a5ec TT |
158 | * stripe size. This should result in better allocation on RAID setups. If |
159 | * not, we search in the specific group using bitmap for best extents. The | |
160 | * tunable min_to_scan and max_to_scan control the behaviour here. | |
c9de560d | 161 | * min_to_scan indicate how long the mballoc __must__ look for a best |
b713a5ec | 162 | * extent and max_to_scan indicates how long the mballoc __can__ look for a |
c9de560d AT |
163 | * best extent in the found extents. Searching for the blocks starts with |
164 | * the group specified as the goal value in allocation context via | |
165 | * ac_g_ex. Each group is first checked based on the criteria whether it | |
caaf7a29 | 166 | * can be used for allocation. ext4_mb_good_group explains how the groups are |
c9de560d AT |
167 | * checked. |
168 | * | |
169 | * Both the prealloc space are getting populated as above. So for the first | |
170 | * request we will hit the buddy cache which will result in this prealloc | |
171 | * space getting filled. The prealloc space is then later used for the | |
172 | * subsequent request. | |
173 | */ | |
174 | ||
175 | /* | |
176 | * mballoc operates on the following data: | |
177 | * - on-disk bitmap | |
178 | * - in-core buddy (actually includes buddy and bitmap) | |
179 | * - preallocation descriptors (PAs) | |
180 | * | |
181 | * there are two types of preallocations: | |
182 | * - inode | |
183 | * assiged to specific inode and can be used for this inode only. | |
184 | * it describes part of inode's space preallocated to specific | |
185 | * physical blocks. any block from that preallocated can be used | |
186 | * independent. the descriptor just tracks number of blocks left | |
187 | * unused. so, before taking some block from descriptor, one must | |
188 | * make sure corresponded logical block isn't allocated yet. this | |
189 | * also means that freeing any block within descriptor's range | |
190 | * must discard all preallocated blocks. | |
191 | * - locality group | |
192 | * assigned to specific locality group which does not translate to | |
193 | * permanent set of inodes: inode can join and leave group. space | |
194 | * from this type of preallocation can be used for any inode. thus | |
195 | * it's consumed from the beginning to the end. | |
196 | * | |
197 | * relation between them can be expressed as: | |
198 | * in-core buddy = on-disk bitmap + preallocation descriptors | |
199 | * | |
200 | * this mean blocks mballoc considers used are: | |
201 | * - allocated blocks (persistent) | |
202 | * - preallocated blocks (non-persistent) | |
203 | * | |
204 | * consistency in mballoc world means that at any time a block is either | |
205 | * free or used in ALL structures. notice: "any time" should not be read | |
206 | * literally -- time is discrete and delimited by locks. | |
207 | * | |
208 | * to keep it simple, we don't use block numbers, instead we count number of | |
209 | * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA. | |
210 | * | |
211 | * all operations can be expressed as: | |
212 | * - init buddy: buddy = on-disk + PAs | |
213 | * - new PA: buddy += N; PA = N | |
214 | * - use inode PA: on-disk += N; PA -= N | |
215 | * - discard inode PA buddy -= on-disk - PA; PA = 0 | |
216 | * - use locality group PA on-disk += N; PA -= N | |
217 | * - discard locality group PA buddy -= PA; PA = 0 | |
218 | * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap | |
219 | * is used in real operation because we can't know actual used | |
220 | * bits from PA, only from on-disk bitmap | |
221 | * | |
222 | * if we follow this strict logic, then all operations above should be atomic. | |
223 | * given some of them can block, we'd have to use something like semaphores | |
224 | * killing performance on high-end SMP hardware. let's try to relax it using | |
225 | * the following knowledge: | |
226 | * 1) if buddy is referenced, it's already initialized | |
227 | * 2) while block is used in buddy and the buddy is referenced, | |
228 | * nobody can re-allocate that block | |
229 | * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has | |
230 | * bit set and PA claims same block, it's OK. IOW, one can set bit in | |
231 | * on-disk bitmap if buddy has same bit set or/and PA covers corresponded | |
232 | * block | |
233 | * | |
234 | * so, now we're building a concurrency table: | |
235 | * - init buddy vs. | |
236 | * - new PA | |
237 | * blocks for PA are allocated in the buddy, buddy must be referenced | |
238 | * until PA is linked to allocation group to avoid concurrent buddy init | |
239 | * - use inode PA | |
240 | * we need to make sure that either on-disk bitmap or PA has uptodate data | |
241 | * given (3) we care that PA-=N operation doesn't interfere with init | |
242 | * - discard inode PA | |
243 | * the simplest way would be to have buddy initialized by the discard | |
244 | * - use locality group PA | |
245 | * again PA-=N must be serialized with init | |
246 | * - discard locality group PA | |
247 | * the simplest way would be to have buddy initialized by the discard | |
248 | * - new PA vs. | |
249 | * - use inode PA | |
250 | * i_data_sem serializes them | |
251 | * - discard inode PA | |
252 | * discard process must wait until PA isn't used by another process | |
253 | * - use locality group PA | |
254 | * some mutex should serialize them | |
255 | * - discard locality group PA | |
256 | * discard process must wait until PA isn't used by another process | |
257 | * - use inode PA | |
258 | * - use inode PA | |
259 | * i_data_sem or another mutex should serializes them | |
260 | * - discard inode PA | |
261 | * discard process must wait until PA isn't used by another process | |
262 | * - use locality group PA | |
263 | * nothing wrong here -- they're different PAs covering different blocks | |
264 | * - discard locality group PA | |
265 | * discard process must wait until PA isn't used by another process | |
266 | * | |
267 | * now we're ready to make few consequences: | |
268 | * - PA is referenced and while it is no discard is possible | |
269 | * - PA is referenced until block isn't marked in on-disk bitmap | |
270 | * - PA changes only after on-disk bitmap | |
271 | * - discard must not compete with init. either init is done before | |
272 | * any discard or they're serialized somehow | |
273 | * - buddy init as sum of on-disk bitmap and PAs is done atomically | |
274 | * | |
275 | * a special case when we've used PA to emptiness. no need to modify buddy | |
276 | * in this case, but we should care about concurrent init | |
277 | * | |
278 | */ | |
279 | ||
280 | /* | |
281 | * Logic in few words: | |
282 | * | |
283 | * - allocation: | |
284 | * load group | |
285 | * find blocks | |
286 | * mark bits in on-disk bitmap | |
287 | * release group | |
288 | * | |
289 | * - use preallocation: | |
290 | * find proper PA (per-inode or group) | |
291 | * load group | |
292 | * mark bits in on-disk bitmap | |
293 | * release group | |
294 | * release PA | |
295 | * | |
296 | * - free: | |
297 | * load group | |
298 | * mark bits in on-disk bitmap | |
299 | * release group | |
300 | * | |
301 | * - discard preallocations in group: | |
302 | * mark PAs deleted | |
303 | * move them onto local list | |
304 | * load on-disk bitmap | |
305 | * load group | |
306 | * remove PA from object (inode or locality group) | |
307 | * mark free blocks in-core | |
308 | * | |
309 | * - discard inode's preallocations: | |
310 | */ | |
311 | ||
312 | /* | |
313 | * Locking rules | |
314 | * | |
315 | * Locks: | |
316 | * - bitlock on a group (group) | |
317 | * - object (inode/locality) (object) | |
318 | * - per-pa lock (pa) | |
319 | * | |
320 | * Paths: | |
321 | * - new pa | |
322 | * object | |
323 | * group | |
324 | * | |
325 | * - find and use pa: | |
326 | * pa | |
327 | * | |
328 | * - release consumed pa: | |
329 | * pa | |
330 | * group | |
331 | * object | |
332 | * | |
333 | * - generate in-core bitmap: | |
334 | * group | |
335 | * pa | |
336 | * | |
337 | * - discard all for given object (inode, locality group): | |
338 | * object | |
339 | * pa | |
340 | * group | |
341 | * | |
342 | * - discard all for given group: | |
343 | * group | |
344 | * pa | |
345 | * group | |
346 | * object | |
347 | * | |
348 | */ | |
c3a326a6 AK |
349 | static struct kmem_cache *ext4_pspace_cachep; |
350 | static struct kmem_cache *ext4_ac_cachep; | |
18aadd47 | 351 | static struct kmem_cache *ext4_free_data_cachep; |
fb1813f4 CW |
352 | |
353 | /* We create slab caches for groupinfo data structures based on the | |
354 | * superblock block size. There will be one per mounted filesystem for | |
355 | * each unique s_blocksize_bits */ | |
2892c15d | 356 | #define NR_GRPINFO_CACHES 8 |
fb1813f4 CW |
357 | static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES]; |
358 | ||
2892c15d ES |
359 | static const char *ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = { |
360 | "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k", | |
361 | "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k", | |
362 | "ext4_groupinfo_64k", "ext4_groupinfo_128k" | |
363 | }; | |
364 | ||
c3a326a6 AK |
365 | static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, |
366 | ext4_group_t group); | |
7a2fcbf7 AK |
367 | static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, |
368 | ext4_group_t group); | |
18aadd47 BJ |
369 | static void ext4_free_data_callback(struct super_block *sb, |
370 | struct ext4_journal_cb_entry *jce, int rc); | |
c3a326a6 | 371 | |
ffad0a44 AK |
372 | static inline void *mb_correct_addr_and_bit(int *bit, void *addr) |
373 | { | |
c9de560d | 374 | #if BITS_PER_LONG == 64 |
ffad0a44 AK |
375 | *bit += ((unsigned long) addr & 7UL) << 3; |
376 | addr = (void *) ((unsigned long) addr & ~7UL); | |
c9de560d | 377 | #elif BITS_PER_LONG == 32 |
ffad0a44 AK |
378 | *bit += ((unsigned long) addr & 3UL) << 3; |
379 | addr = (void *) ((unsigned long) addr & ~3UL); | |
c9de560d AT |
380 | #else |
381 | #error "how many bits you are?!" | |
382 | #endif | |
ffad0a44 AK |
383 | return addr; |
384 | } | |
c9de560d AT |
385 | |
386 | static inline int mb_test_bit(int bit, void *addr) | |
387 | { | |
388 | /* | |
389 | * ext4_test_bit on architecture like powerpc | |
390 | * needs unsigned long aligned address | |
391 | */ | |
ffad0a44 | 392 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
393 | return ext4_test_bit(bit, addr); |
394 | } | |
395 | ||
396 | static inline void mb_set_bit(int bit, void *addr) | |
397 | { | |
ffad0a44 | 398 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
399 | ext4_set_bit(bit, addr); |
400 | } | |
401 | ||
c9de560d AT |
402 | static inline void mb_clear_bit(int bit, void *addr) |
403 | { | |
ffad0a44 | 404 | addr = mb_correct_addr_and_bit(&bit, addr); |
c9de560d AT |
405 | ext4_clear_bit(bit, addr); |
406 | } | |
407 | ||
eabe0444 AS |
408 | static inline int mb_test_and_clear_bit(int bit, void *addr) |
409 | { | |
410 | addr = mb_correct_addr_and_bit(&bit, addr); | |
411 | return ext4_test_and_clear_bit(bit, addr); | |
412 | } | |
413 | ||
ffad0a44 AK |
414 | static inline int mb_find_next_zero_bit(void *addr, int max, int start) |
415 | { | |
e7dfb246 | 416 | int fix = 0, ret, tmpmax; |
ffad0a44 | 417 | addr = mb_correct_addr_and_bit(&fix, addr); |
e7dfb246 | 418 | tmpmax = max + fix; |
ffad0a44 AK |
419 | start += fix; |
420 | ||
e7dfb246 AK |
421 | ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix; |
422 | if (ret > max) | |
423 | return max; | |
424 | return ret; | |
ffad0a44 AK |
425 | } |
426 | ||
427 | static inline int mb_find_next_bit(void *addr, int max, int start) | |
428 | { | |
e7dfb246 | 429 | int fix = 0, ret, tmpmax; |
ffad0a44 | 430 | addr = mb_correct_addr_and_bit(&fix, addr); |
e7dfb246 | 431 | tmpmax = max + fix; |
ffad0a44 AK |
432 | start += fix; |
433 | ||
e7dfb246 AK |
434 | ret = ext4_find_next_bit(addr, tmpmax, start) - fix; |
435 | if (ret > max) | |
436 | return max; | |
437 | return ret; | |
ffad0a44 AK |
438 | } |
439 | ||
c9de560d AT |
440 | static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max) |
441 | { | |
442 | char *bb; | |
443 | ||
c5e8f3f3 | 444 | BUG_ON(e4b->bd_bitmap == e4b->bd_buddy); |
c9de560d AT |
445 | BUG_ON(max == NULL); |
446 | ||
447 | if (order > e4b->bd_blkbits + 1) { | |
448 | *max = 0; | |
449 | return NULL; | |
450 | } | |
451 | ||
452 | /* at order 0 we see each particular block */ | |
84b775a3 CL |
453 | if (order == 0) { |
454 | *max = 1 << (e4b->bd_blkbits + 3); | |
c5e8f3f3 | 455 | return e4b->bd_bitmap; |
84b775a3 | 456 | } |
c9de560d | 457 | |
c5e8f3f3 | 458 | bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order]; |
c9de560d AT |
459 | *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order]; |
460 | ||
461 | return bb; | |
462 | } | |
463 | ||
464 | #ifdef DOUBLE_CHECK | |
465 | static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b, | |
466 | int first, int count) | |
467 | { | |
468 | int i; | |
469 | struct super_block *sb = e4b->bd_sb; | |
470 | ||
471 | if (unlikely(e4b->bd_info->bb_bitmap == NULL)) | |
472 | return; | |
bc8e6740 | 473 | assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); |
c9de560d AT |
474 | for (i = 0; i < count; i++) { |
475 | if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) { | |
476 | ext4_fsblk_t blocknr; | |
5661bd68 AM |
477 | |
478 | blocknr = ext4_group_first_block_no(sb, e4b->bd_group); | |
53accfa9 | 479 | blocknr += EXT4_C2B(EXT4_SB(sb), first + i); |
5d1b1b3f | 480 | ext4_grp_locked_error(sb, e4b->bd_group, |
e29136f8 TT |
481 | inode ? inode->i_ino : 0, |
482 | blocknr, | |
483 | "freeing block already freed " | |
484 | "(bit %u)", | |
485 | first + i); | |
c9de560d AT |
486 | } |
487 | mb_clear_bit(first + i, e4b->bd_info->bb_bitmap); | |
488 | } | |
489 | } | |
490 | ||
491 | static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count) | |
492 | { | |
493 | int i; | |
494 | ||
495 | if (unlikely(e4b->bd_info->bb_bitmap == NULL)) | |
496 | return; | |
bc8e6740 | 497 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
498 | for (i = 0; i < count; i++) { |
499 | BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap)); | |
500 | mb_set_bit(first + i, e4b->bd_info->bb_bitmap); | |
501 | } | |
502 | } | |
503 | ||
504 | static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) | |
505 | { | |
506 | if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) { | |
507 | unsigned char *b1, *b2; | |
508 | int i; | |
509 | b1 = (unsigned char *) e4b->bd_info->bb_bitmap; | |
510 | b2 = (unsigned char *) bitmap; | |
511 | for (i = 0; i < e4b->bd_sb->s_blocksize; i++) { | |
512 | if (b1[i] != b2[i]) { | |
9d8b9ec4 TT |
513 | ext4_msg(e4b->bd_sb, KERN_ERR, |
514 | "corruption in group %u " | |
515 | "at byte %u(%u): %x in copy != %x " | |
516 | "on disk/prealloc", | |
517 | e4b->bd_group, i, i * 8, b1[i], b2[i]); | |
c9de560d AT |
518 | BUG(); |
519 | } | |
520 | } | |
521 | } | |
522 | } | |
523 | ||
524 | #else | |
525 | static inline void mb_free_blocks_double(struct inode *inode, | |
526 | struct ext4_buddy *e4b, int first, int count) | |
527 | { | |
528 | return; | |
529 | } | |
530 | static inline void mb_mark_used_double(struct ext4_buddy *e4b, | |
531 | int first, int count) | |
532 | { | |
533 | return; | |
534 | } | |
535 | static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) | |
536 | { | |
537 | return; | |
538 | } | |
539 | #endif | |
540 | ||
541 | #ifdef AGGRESSIVE_CHECK | |
542 | ||
543 | #define MB_CHECK_ASSERT(assert) \ | |
544 | do { \ | |
545 | if (!(assert)) { \ | |
546 | printk(KERN_EMERG \ | |
547 | "Assertion failure in %s() at %s:%d: \"%s\"\n", \ | |
548 | function, file, line, # assert); \ | |
549 | BUG(); \ | |
550 | } \ | |
551 | } while (0) | |
552 | ||
553 | static int __mb_check_buddy(struct ext4_buddy *e4b, char *file, | |
554 | const char *function, int line) | |
555 | { | |
556 | struct super_block *sb = e4b->bd_sb; | |
557 | int order = e4b->bd_blkbits + 1; | |
558 | int max; | |
559 | int max2; | |
560 | int i; | |
561 | int j; | |
562 | int k; | |
563 | int count; | |
564 | struct ext4_group_info *grp; | |
565 | int fragments = 0; | |
566 | int fstart; | |
567 | struct list_head *cur; | |
568 | void *buddy; | |
569 | void *buddy2; | |
570 | ||
c9de560d AT |
571 | { |
572 | static int mb_check_counter; | |
573 | if (mb_check_counter++ % 100 != 0) | |
574 | return 0; | |
575 | } | |
576 | ||
577 | while (order > 1) { | |
578 | buddy = mb_find_buddy(e4b, order, &max); | |
579 | MB_CHECK_ASSERT(buddy); | |
580 | buddy2 = mb_find_buddy(e4b, order - 1, &max2); | |
581 | MB_CHECK_ASSERT(buddy2); | |
582 | MB_CHECK_ASSERT(buddy != buddy2); | |
583 | MB_CHECK_ASSERT(max * 2 == max2); | |
584 | ||
585 | count = 0; | |
586 | for (i = 0; i < max; i++) { | |
587 | ||
588 | if (mb_test_bit(i, buddy)) { | |
589 | /* only single bit in buddy2 may be 1 */ | |
590 | if (!mb_test_bit(i << 1, buddy2)) { | |
591 | MB_CHECK_ASSERT( | |
592 | mb_test_bit((i<<1)+1, buddy2)); | |
593 | } else if (!mb_test_bit((i << 1) + 1, buddy2)) { | |
594 | MB_CHECK_ASSERT( | |
595 | mb_test_bit(i << 1, buddy2)); | |
596 | } | |
597 | continue; | |
598 | } | |
599 | ||
0a10da73 | 600 | /* both bits in buddy2 must be 1 */ |
c9de560d AT |
601 | MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2)); |
602 | MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2)); | |
603 | ||
604 | for (j = 0; j < (1 << order); j++) { | |
605 | k = (i * (1 << order)) + j; | |
606 | MB_CHECK_ASSERT( | |
c5e8f3f3 | 607 | !mb_test_bit(k, e4b->bd_bitmap)); |
c9de560d AT |
608 | } |
609 | count++; | |
610 | } | |
611 | MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count); | |
612 | order--; | |
613 | } | |
614 | ||
615 | fstart = -1; | |
616 | buddy = mb_find_buddy(e4b, 0, &max); | |
617 | for (i = 0; i < max; i++) { | |
618 | if (!mb_test_bit(i, buddy)) { | |
619 | MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free); | |
620 | if (fstart == -1) { | |
621 | fragments++; | |
622 | fstart = i; | |
623 | } | |
624 | continue; | |
625 | } | |
626 | fstart = -1; | |
627 | /* check used bits only */ | |
628 | for (j = 0; j < e4b->bd_blkbits + 1; j++) { | |
629 | buddy2 = mb_find_buddy(e4b, j, &max2); | |
630 | k = i >> j; | |
631 | MB_CHECK_ASSERT(k < max2); | |
632 | MB_CHECK_ASSERT(mb_test_bit(k, buddy2)); | |
633 | } | |
634 | } | |
635 | MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info)); | |
636 | MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments); | |
637 | ||
638 | grp = ext4_get_group_info(sb, e4b->bd_group); | |
c9de560d AT |
639 | list_for_each(cur, &grp->bb_prealloc_list) { |
640 | ext4_group_t groupnr; | |
641 | struct ext4_prealloc_space *pa; | |
60bd63d1 SR |
642 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); |
643 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k); | |
c9de560d | 644 | MB_CHECK_ASSERT(groupnr == e4b->bd_group); |
60bd63d1 | 645 | for (i = 0; i < pa->pa_len; i++) |
c9de560d AT |
646 | MB_CHECK_ASSERT(mb_test_bit(k + i, buddy)); |
647 | } | |
648 | return 0; | |
649 | } | |
650 | #undef MB_CHECK_ASSERT | |
651 | #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \ | |
46e665e9 | 652 | __FILE__, __func__, __LINE__) |
c9de560d AT |
653 | #else |
654 | #define mb_check_buddy(e4b) | |
655 | #endif | |
656 | ||
7c786059 CL |
657 | /* |
658 | * Divide blocks started from @first with length @len into | |
659 | * smaller chunks with power of 2 blocks. | |
660 | * Clear the bits in bitmap which the blocks of the chunk(s) covered, | |
661 | * then increase bb_counters[] for corresponded chunk size. | |
662 | */ | |
c9de560d | 663 | static void ext4_mb_mark_free_simple(struct super_block *sb, |
a36b4498 | 664 | void *buddy, ext4_grpblk_t first, ext4_grpblk_t len, |
c9de560d AT |
665 | struct ext4_group_info *grp) |
666 | { | |
667 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
a36b4498 ES |
668 | ext4_grpblk_t min; |
669 | ext4_grpblk_t max; | |
670 | ext4_grpblk_t chunk; | |
c9de560d AT |
671 | unsigned short border; |
672 | ||
7137d7a4 | 673 | BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb)); |
c9de560d AT |
674 | |
675 | border = 2 << sb->s_blocksize_bits; | |
676 | ||
677 | while (len > 0) { | |
678 | /* find how many blocks can be covered since this position */ | |
679 | max = ffs(first | border) - 1; | |
680 | ||
681 | /* find how many blocks of power 2 we need to mark */ | |
682 | min = fls(len) - 1; | |
683 | ||
684 | if (max < min) | |
685 | min = max; | |
686 | chunk = 1 << min; | |
687 | ||
688 | /* mark multiblock chunks only */ | |
689 | grp->bb_counters[min]++; | |
690 | if (min > 0) | |
691 | mb_clear_bit(first >> min, | |
692 | buddy + sbi->s_mb_offsets[min]); | |
693 | ||
694 | len -= chunk; | |
695 | first += chunk; | |
696 | } | |
697 | } | |
698 | ||
8a57d9d6 CW |
699 | /* |
700 | * Cache the order of the largest free extent we have available in this block | |
701 | * group. | |
702 | */ | |
703 | static void | |
704 | mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp) | |
705 | { | |
706 | int i; | |
707 | int bits; | |
708 | ||
709 | grp->bb_largest_free_order = -1; /* uninit */ | |
710 | ||
711 | bits = sb->s_blocksize_bits + 1; | |
712 | for (i = bits; i >= 0; i--) { | |
713 | if (grp->bb_counters[i] > 0) { | |
714 | grp->bb_largest_free_order = i; | |
715 | break; | |
716 | } | |
717 | } | |
718 | } | |
719 | ||
089ceecc ES |
720 | static noinline_for_stack |
721 | void ext4_mb_generate_buddy(struct super_block *sb, | |
c9de560d AT |
722 | void *buddy, void *bitmap, ext4_group_t group) |
723 | { | |
724 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
e43bb4e6 | 725 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
7137d7a4 | 726 | ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb); |
a36b4498 ES |
727 | ext4_grpblk_t i = 0; |
728 | ext4_grpblk_t first; | |
729 | ext4_grpblk_t len; | |
c9de560d AT |
730 | unsigned free = 0; |
731 | unsigned fragments = 0; | |
732 | unsigned long long period = get_cycles(); | |
733 | ||
734 | /* initialize buddy from bitmap which is aggregation | |
735 | * of on-disk bitmap and preallocations */ | |
ffad0a44 | 736 | i = mb_find_next_zero_bit(bitmap, max, 0); |
c9de560d AT |
737 | grp->bb_first_free = i; |
738 | while (i < max) { | |
739 | fragments++; | |
740 | first = i; | |
ffad0a44 | 741 | i = mb_find_next_bit(bitmap, max, i); |
c9de560d AT |
742 | len = i - first; |
743 | free += len; | |
744 | if (len > 1) | |
745 | ext4_mb_mark_free_simple(sb, buddy, first, len, grp); | |
746 | else | |
747 | grp->bb_counters[0]++; | |
748 | if (i < max) | |
ffad0a44 | 749 | i = mb_find_next_zero_bit(bitmap, max, i); |
c9de560d AT |
750 | } |
751 | grp->bb_fragments = fragments; | |
752 | ||
753 | if (free != grp->bb_free) { | |
e29136f8 | 754 | ext4_grp_locked_error(sb, group, 0, 0, |
94d4c066 TT |
755 | "block bitmap and bg descriptor " |
756 | "inconsistent: %u vs %u free clusters", | |
e29136f8 | 757 | free, grp->bb_free); |
e56eb659 | 758 | /* |
163a203d | 759 | * If we intend to continue, we consider group descriptor |
e56eb659 AK |
760 | * corrupt and update bb_free using bitmap value |
761 | */ | |
c9de560d | 762 | grp->bb_free = free; |
e43bb4e6 NJ |
763 | if (!EXT4_MB_GRP_BBITMAP_CORRUPT(grp)) |
764 | percpu_counter_sub(&sbi->s_freeclusters_counter, | |
765 | grp->bb_free); | |
163a203d | 766 | set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT, &grp->bb_state); |
c9de560d | 767 | } |
8a57d9d6 | 768 | mb_set_largest_free_order(sb, grp); |
c9de560d AT |
769 | |
770 | clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state)); | |
771 | ||
772 | period = get_cycles() - period; | |
773 | spin_lock(&EXT4_SB(sb)->s_bal_lock); | |
774 | EXT4_SB(sb)->s_mb_buddies_generated++; | |
775 | EXT4_SB(sb)->s_mb_generation_time += period; | |
776 | spin_unlock(&EXT4_SB(sb)->s_bal_lock); | |
777 | } | |
778 | ||
eabe0444 AS |
779 | static void mb_regenerate_buddy(struct ext4_buddy *e4b) |
780 | { | |
781 | int count; | |
782 | int order = 1; | |
783 | void *buddy; | |
784 | ||
785 | while ((buddy = mb_find_buddy(e4b, order++, &count))) { | |
786 | ext4_set_bits(buddy, 0, count); | |
787 | } | |
788 | e4b->bd_info->bb_fragments = 0; | |
789 | memset(e4b->bd_info->bb_counters, 0, | |
790 | sizeof(*e4b->bd_info->bb_counters) * | |
791 | (e4b->bd_sb->s_blocksize_bits + 2)); | |
792 | ||
793 | ext4_mb_generate_buddy(e4b->bd_sb, e4b->bd_buddy, | |
794 | e4b->bd_bitmap, e4b->bd_group); | |
795 | } | |
796 | ||
c9de560d AT |
797 | /* The buddy information is attached the buddy cache inode |
798 | * for convenience. The information regarding each group | |
799 | * is loaded via ext4_mb_load_buddy. The information involve | |
800 | * block bitmap and buddy information. The information are | |
801 | * stored in the inode as | |
802 | * | |
803 | * { page } | |
c3a326a6 | 804 | * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... |
c9de560d AT |
805 | * |
806 | * | |
807 | * one block each for bitmap and buddy information. | |
808 | * So for each group we take up 2 blocks. A page can | |
809 | * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks. | |
810 | * So it can have information regarding groups_per_page which | |
811 | * is blocks_per_page/2 | |
8a57d9d6 CW |
812 | * |
813 | * Locking note: This routine takes the block group lock of all groups | |
814 | * for this page; do not hold this lock when calling this routine! | |
c9de560d AT |
815 | */ |
816 | ||
817 | static int ext4_mb_init_cache(struct page *page, char *incore) | |
818 | { | |
8df9675f | 819 | ext4_group_t ngroups; |
c9de560d AT |
820 | int blocksize; |
821 | int blocks_per_page; | |
822 | int groups_per_page; | |
823 | int err = 0; | |
824 | int i; | |
813e5727 | 825 | ext4_group_t first_group, group; |
c9de560d AT |
826 | int first_block; |
827 | struct super_block *sb; | |
828 | struct buffer_head *bhs; | |
fa77dcfa | 829 | struct buffer_head **bh = NULL; |
c9de560d AT |
830 | struct inode *inode; |
831 | char *data; | |
832 | char *bitmap; | |
9b8b7d35 | 833 | struct ext4_group_info *grinfo; |
c9de560d | 834 | |
6ba495e9 | 835 | mb_debug(1, "init page %lu\n", page->index); |
c9de560d AT |
836 | |
837 | inode = page->mapping->host; | |
838 | sb = inode->i_sb; | |
8df9675f | 839 | ngroups = ext4_get_groups_count(sb); |
c9de560d AT |
840 | blocksize = 1 << inode->i_blkbits; |
841 | blocks_per_page = PAGE_CACHE_SIZE / blocksize; | |
842 | ||
843 | groups_per_page = blocks_per_page >> 1; | |
844 | if (groups_per_page == 0) | |
845 | groups_per_page = 1; | |
846 | ||
847 | /* allocate buffer_heads to read bitmaps */ | |
848 | if (groups_per_page > 1) { | |
c9de560d AT |
849 | i = sizeof(struct buffer_head *) * groups_per_page; |
850 | bh = kzalloc(i, GFP_NOFS); | |
813e5727 TT |
851 | if (bh == NULL) { |
852 | err = -ENOMEM; | |
c9de560d | 853 | goto out; |
813e5727 | 854 | } |
c9de560d AT |
855 | } else |
856 | bh = &bhs; | |
857 | ||
858 | first_group = page->index * blocks_per_page / 2; | |
859 | ||
860 | /* read all groups the page covers into the cache */ | |
813e5727 TT |
861 | for (i = 0, group = first_group; i < groups_per_page; i++, group++) { |
862 | if (group >= ngroups) | |
c9de560d AT |
863 | break; |
864 | ||
813e5727 | 865 | grinfo = ext4_get_group_info(sb, group); |
9b8b7d35 AG |
866 | /* |
867 | * If page is uptodate then we came here after online resize | |
868 | * which added some new uninitialized group info structs, so | |
869 | * we must skip all initialized uptodate buddies on the page, | |
870 | * which may be currently in use by an allocating task. | |
871 | */ | |
872 | if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) { | |
873 | bh[i] = NULL; | |
874 | continue; | |
875 | } | |
813e5727 TT |
876 | if (!(bh[i] = ext4_read_block_bitmap_nowait(sb, group))) { |
877 | err = -ENOMEM; | |
c9de560d | 878 | goto out; |
2ccb5fb9 | 879 | } |
813e5727 | 880 | mb_debug(1, "read bitmap for group %u\n", group); |
c9de560d AT |
881 | } |
882 | ||
883 | /* wait for I/O completion */ | |
813e5727 TT |
884 | for (i = 0, group = first_group; i < groups_per_page; i++, group++) { |
885 | if (bh[i] && ext4_wait_block_bitmap(sb, group, bh[i])) { | |
886 | err = -EIO; | |
c9de560d | 887 | goto out; |
813e5727 TT |
888 | } |
889 | } | |
c9de560d AT |
890 | |
891 | first_block = page->index * blocks_per_page; | |
892 | for (i = 0; i < blocks_per_page; i++) { | |
c9de560d | 893 | group = (first_block + i) >> 1; |
8df9675f | 894 | if (group >= ngroups) |
c9de560d AT |
895 | break; |
896 | ||
9b8b7d35 AG |
897 | if (!bh[group - first_group]) |
898 | /* skip initialized uptodate buddy */ | |
899 | continue; | |
900 | ||
c9de560d AT |
901 | /* |
902 | * data carry information regarding this | |
903 | * particular group in the format specified | |
904 | * above | |
905 | * | |
906 | */ | |
907 | data = page_address(page) + (i * blocksize); | |
908 | bitmap = bh[group - first_group]->b_data; | |
909 | ||
910 | /* | |
911 | * We place the buddy block and bitmap block | |
912 | * close together | |
913 | */ | |
914 | if ((first_block + i) & 1) { | |
915 | /* this is block of buddy */ | |
916 | BUG_ON(incore == NULL); | |
6ba495e9 | 917 | mb_debug(1, "put buddy for group %u in page %lu/%x\n", |
c9de560d | 918 | group, page->index, i * blocksize); |
f307333e | 919 | trace_ext4_mb_buddy_bitmap_load(sb, group); |
c9de560d AT |
920 | grinfo = ext4_get_group_info(sb, group); |
921 | grinfo->bb_fragments = 0; | |
922 | memset(grinfo->bb_counters, 0, | |
1927805e ES |
923 | sizeof(*grinfo->bb_counters) * |
924 | (sb->s_blocksize_bits+2)); | |
c9de560d AT |
925 | /* |
926 | * incore got set to the group block bitmap below | |
927 | */ | |
7a2fcbf7 | 928 | ext4_lock_group(sb, group); |
9b8b7d35 AG |
929 | /* init the buddy */ |
930 | memset(data, 0xff, blocksize); | |
c9de560d | 931 | ext4_mb_generate_buddy(sb, data, incore, group); |
7a2fcbf7 | 932 | ext4_unlock_group(sb, group); |
c9de560d AT |
933 | incore = NULL; |
934 | } else { | |
935 | /* this is block of bitmap */ | |
936 | BUG_ON(incore != NULL); | |
6ba495e9 | 937 | mb_debug(1, "put bitmap for group %u in page %lu/%x\n", |
c9de560d | 938 | group, page->index, i * blocksize); |
f307333e | 939 | trace_ext4_mb_bitmap_load(sb, group); |
c9de560d AT |
940 | |
941 | /* see comments in ext4_mb_put_pa() */ | |
942 | ext4_lock_group(sb, group); | |
943 | memcpy(data, bitmap, blocksize); | |
944 | ||
945 | /* mark all preallocated blks used in in-core bitmap */ | |
946 | ext4_mb_generate_from_pa(sb, data, group); | |
7a2fcbf7 | 947 | ext4_mb_generate_from_freelist(sb, data, group); |
c9de560d AT |
948 | ext4_unlock_group(sb, group); |
949 | ||
950 | /* set incore so that the buddy information can be | |
951 | * generated using this | |
952 | */ | |
953 | incore = data; | |
954 | } | |
955 | } | |
956 | SetPageUptodate(page); | |
957 | ||
958 | out: | |
959 | if (bh) { | |
9b8b7d35 | 960 | for (i = 0; i < groups_per_page; i++) |
c9de560d AT |
961 | brelse(bh[i]); |
962 | if (bh != &bhs) | |
963 | kfree(bh); | |
964 | } | |
965 | return err; | |
966 | } | |
967 | ||
eee4adc7 | 968 | /* |
2de8807b AG |
969 | * Lock the buddy and bitmap pages. This make sure other parallel init_group |
970 | * on the same buddy page doesn't happen whild holding the buddy page lock. | |
971 | * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap | |
972 | * are on the same page e4b->bd_buddy_page is NULL and return value is 0. | |
eee4adc7 | 973 | */ |
2de8807b AG |
974 | static int ext4_mb_get_buddy_page_lock(struct super_block *sb, |
975 | ext4_group_t group, struct ext4_buddy *e4b) | |
eee4adc7 | 976 | { |
2de8807b AG |
977 | struct inode *inode = EXT4_SB(sb)->s_buddy_cache; |
978 | int block, pnum, poff; | |
eee4adc7 | 979 | int blocks_per_page; |
2de8807b AG |
980 | struct page *page; |
981 | ||
982 | e4b->bd_buddy_page = NULL; | |
983 | e4b->bd_bitmap_page = NULL; | |
eee4adc7 ES |
984 | |
985 | blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; | |
986 | /* | |
987 | * the buddy cache inode stores the block bitmap | |
988 | * and buddy information in consecutive blocks. | |
989 | * So for each group we need two blocks. | |
990 | */ | |
991 | block = group * 2; | |
992 | pnum = block / blocks_per_page; | |
2de8807b AG |
993 | poff = block % blocks_per_page; |
994 | page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); | |
995 | if (!page) | |
c57ab39b | 996 | return -ENOMEM; |
2de8807b AG |
997 | BUG_ON(page->mapping != inode->i_mapping); |
998 | e4b->bd_bitmap_page = page; | |
999 | e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); | |
1000 | ||
1001 | if (blocks_per_page >= 2) { | |
1002 | /* buddy and bitmap are on the same page */ | |
1003 | return 0; | |
eee4adc7 | 1004 | } |
2de8807b AG |
1005 | |
1006 | block++; | |
1007 | pnum = block / blocks_per_page; | |
2de8807b AG |
1008 | page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); |
1009 | if (!page) | |
c57ab39b | 1010 | return -ENOMEM; |
2de8807b AG |
1011 | BUG_ON(page->mapping != inode->i_mapping); |
1012 | e4b->bd_buddy_page = page; | |
1013 | return 0; | |
eee4adc7 ES |
1014 | } |
1015 | ||
2de8807b | 1016 | static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b) |
eee4adc7 | 1017 | { |
2de8807b AG |
1018 | if (e4b->bd_bitmap_page) { |
1019 | unlock_page(e4b->bd_bitmap_page); | |
1020 | page_cache_release(e4b->bd_bitmap_page); | |
1021 | } | |
1022 | if (e4b->bd_buddy_page) { | |
1023 | unlock_page(e4b->bd_buddy_page); | |
1024 | page_cache_release(e4b->bd_buddy_page); | |
eee4adc7 | 1025 | } |
eee4adc7 ES |
1026 | } |
1027 | ||
8a57d9d6 CW |
1028 | /* |
1029 | * Locking note: This routine calls ext4_mb_init_cache(), which takes the | |
1030 | * block group lock of all groups for this page; do not hold the BG lock when | |
1031 | * calling this routine! | |
1032 | */ | |
b6a758ec AK |
1033 | static noinline_for_stack |
1034 | int ext4_mb_init_group(struct super_block *sb, ext4_group_t group) | |
1035 | { | |
1036 | ||
b6a758ec | 1037 | struct ext4_group_info *this_grp; |
2de8807b AG |
1038 | struct ext4_buddy e4b; |
1039 | struct page *page; | |
1040 | int ret = 0; | |
b6a758ec | 1041 | |
b10a44c3 | 1042 | might_sleep(); |
b6a758ec | 1043 | mb_debug(1, "init group %u\n", group); |
b6a758ec AK |
1044 | this_grp = ext4_get_group_info(sb, group); |
1045 | /* | |
08c3a813 AK |
1046 | * This ensures that we don't reinit the buddy cache |
1047 | * page which map to the group from which we are already | |
1048 | * allocating. If we are looking at the buddy cache we would | |
1049 | * have taken a reference using ext4_mb_load_buddy and that | |
2de8807b | 1050 | * would have pinned buddy page to page cache. |
2457aec6 MG |
1051 | * The call to ext4_mb_get_buddy_page_lock will mark the |
1052 | * page accessed. | |
b6a758ec | 1053 | */ |
2de8807b AG |
1054 | ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b); |
1055 | if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) { | |
b6a758ec AK |
1056 | /* |
1057 | * somebody initialized the group | |
1058 | * return without doing anything | |
1059 | */ | |
b6a758ec AK |
1060 | goto err; |
1061 | } | |
2de8807b AG |
1062 | |
1063 | page = e4b.bd_bitmap_page; | |
1064 | ret = ext4_mb_init_cache(page, NULL); | |
1065 | if (ret) | |
1066 | goto err; | |
1067 | if (!PageUptodate(page)) { | |
b6a758ec AK |
1068 | ret = -EIO; |
1069 | goto err; | |
1070 | } | |
b6a758ec | 1071 | |
2de8807b | 1072 | if (e4b.bd_buddy_page == NULL) { |
b6a758ec AK |
1073 | /* |
1074 | * If both the bitmap and buddy are in | |
1075 | * the same page we don't need to force | |
1076 | * init the buddy | |
1077 | */ | |
2de8807b AG |
1078 | ret = 0; |
1079 | goto err; | |
b6a758ec | 1080 | } |
2de8807b AG |
1081 | /* init buddy cache */ |
1082 | page = e4b.bd_buddy_page; | |
1083 | ret = ext4_mb_init_cache(page, e4b.bd_bitmap); | |
1084 | if (ret) | |
1085 | goto err; | |
1086 | if (!PageUptodate(page)) { | |
b6a758ec AK |
1087 | ret = -EIO; |
1088 | goto err; | |
1089 | } | |
b6a758ec | 1090 | err: |
2de8807b | 1091 | ext4_mb_put_buddy_page_lock(&e4b); |
b6a758ec AK |
1092 | return ret; |
1093 | } | |
1094 | ||
8a57d9d6 CW |
1095 | /* |
1096 | * Locking note: This routine calls ext4_mb_init_cache(), which takes the | |
1097 | * block group lock of all groups for this page; do not hold the BG lock when | |
1098 | * calling this routine! | |
1099 | */ | |
4ddfef7b ES |
1100 | static noinline_for_stack int |
1101 | ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group, | |
1102 | struct ext4_buddy *e4b) | |
c9de560d | 1103 | { |
c9de560d AT |
1104 | int blocks_per_page; |
1105 | int block; | |
1106 | int pnum; | |
1107 | int poff; | |
1108 | struct page *page; | |
fdf6c7a7 | 1109 | int ret; |
920313a7 AK |
1110 | struct ext4_group_info *grp; |
1111 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
1112 | struct inode *inode = sbi->s_buddy_cache; | |
c9de560d | 1113 | |
b10a44c3 | 1114 | might_sleep(); |
6ba495e9 | 1115 | mb_debug(1, "load group %u\n", group); |
c9de560d AT |
1116 | |
1117 | blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; | |
920313a7 | 1118 | grp = ext4_get_group_info(sb, group); |
c9de560d AT |
1119 | |
1120 | e4b->bd_blkbits = sb->s_blocksize_bits; | |
529da704 | 1121 | e4b->bd_info = grp; |
c9de560d AT |
1122 | e4b->bd_sb = sb; |
1123 | e4b->bd_group = group; | |
1124 | e4b->bd_buddy_page = NULL; | |
1125 | e4b->bd_bitmap_page = NULL; | |
1126 | ||
f41c0750 | 1127 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { |
f41c0750 AK |
1128 | /* |
1129 | * we need full data about the group | |
1130 | * to make a good selection | |
1131 | */ | |
1132 | ret = ext4_mb_init_group(sb, group); | |
1133 | if (ret) | |
1134 | return ret; | |
f41c0750 AK |
1135 | } |
1136 | ||
c9de560d AT |
1137 | /* |
1138 | * the buddy cache inode stores the block bitmap | |
1139 | * and buddy information in consecutive blocks. | |
1140 | * So for each group we need two blocks. | |
1141 | */ | |
1142 | block = group * 2; | |
1143 | pnum = block / blocks_per_page; | |
1144 | poff = block % blocks_per_page; | |
1145 | ||
1146 | /* we could use find_or_create_page(), but it locks page | |
1147 | * what we'd like to avoid in fast path ... */ | |
2457aec6 | 1148 | page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED); |
c9de560d AT |
1149 | if (page == NULL || !PageUptodate(page)) { |
1150 | if (page) | |
920313a7 AK |
1151 | /* |
1152 | * drop the page reference and try | |
1153 | * to get the page with lock. If we | |
1154 | * are not uptodate that implies | |
1155 | * somebody just created the page but | |
1156 | * is yet to initialize the same. So | |
1157 | * wait for it to initialize. | |
1158 | */ | |
c9de560d AT |
1159 | page_cache_release(page); |
1160 | page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); | |
1161 | if (page) { | |
1162 | BUG_ON(page->mapping != inode->i_mapping); | |
1163 | if (!PageUptodate(page)) { | |
fdf6c7a7 SF |
1164 | ret = ext4_mb_init_cache(page, NULL); |
1165 | if (ret) { | |
1166 | unlock_page(page); | |
1167 | goto err; | |
1168 | } | |
c9de560d AT |
1169 | mb_cmp_bitmaps(e4b, page_address(page) + |
1170 | (poff * sb->s_blocksize)); | |
1171 | } | |
1172 | unlock_page(page); | |
1173 | } | |
1174 | } | |
c57ab39b YL |
1175 | if (page == NULL) { |
1176 | ret = -ENOMEM; | |
1177 | goto err; | |
1178 | } | |
1179 | if (!PageUptodate(page)) { | |
fdf6c7a7 | 1180 | ret = -EIO; |
c9de560d | 1181 | goto err; |
fdf6c7a7 | 1182 | } |
2457aec6 MG |
1183 | |
1184 | /* Pages marked accessed already */ | |
c9de560d AT |
1185 | e4b->bd_bitmap_page = page; |
1186 | e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); | |
c9de560d AT |
1187 | |
1188 | block++; | |
1189 | pnum = block / blocks_per_page; | |
1190 | poff = block % blocks_per_page; | |
1191 | ||
2457aec6 | 1192 | page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED); |
c9de560d AT |
1193 | if (page == NULL || !PageUptodate(page)) { |
1194 | if (page) | |
1195 | page_cache_release(page); | |
1196 | page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); | |
1197 | if (page) { | |
1198 | BUG_ON(page->mapping != inode->i_mapping); | |
fdf6c7a7 SF |
1199 | if (!PageUptodate(page)) { |
1200 | ret = ext4_mb_init_cache(page, e4b->bd_bitmap); | |
1201 | if (ret) { | |
1202 | unlock_page(page); | |
1203 | goto err; | |
1204 | } | |
1205 | } | |
c9de560d AT |
1206 | unlock_page(page); |
1207 | } | |
1208 | } | |
c57ab39b YL |
1209 | if (page == NULL) { |
1210 | ret = -ENOMEM; | |
1211 | goto err; | |
1212 | } | |
1213 | if (!PageUptodate(page)) { | |
fdf6c7a7 | 1214 | ret = -EIO; |
c9de560d | 1215 | goto err; |
fdf6c7a7 | 1216 | } |
2457aec6 MG |
1217 | |
1218 | /* Pages marked accessed already */ | |
c9de560d AT |
1219 | e4b->bd_buddy_page = page; |
1220 | e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize); | |
c9de560d AT |
1221 | |
1222 | BUG_ON(e4b->bd_bitmap_page == NULL); | |
1223 | BUG_ON(e4b->bd_buddy_page == NULL); | |
1224 | ||
1225 | return 0; | |
1226 | ||
1227 | err: | |
26626f11 YR |
1228 | if (page) |
1229 | page_cache_release(page); | |
c9de560d AT |
1230 | if (e4b->bd_bitmap_page) |
1231 | page_cache_release(e4b->bd_bitmap_page); | |
1232 | if (e4b->bd_buddy_page) | |
1233 | page_cache_release(e4b->bd_buddy_page); | |
1234 | e4b->bd_buddy = NULL; | |
1235 | e4b->bd_bitmap = NULL; | |
fdf6c7a7 | 1236 | return ret; |
c9de560d AT |
1237 | } |
1238 | ||
e39e07fd | 1239 | static void ext4_mb_unload_buddy(struct ext4_buddy *e4b) |
c9de560d AT |
1240 | { |
1241 | if (e4b->bd_bitmap_page) | |
1242 | page_cache_release(e4b->bd_bitmap_page); | |
1243 | if (e4b->bd_buddy_page) | |
1244 | page_cache_release(e4b->bd_buddy_page); | |
1245 | } | |
1246 | ||
1247 | ||
1248 | static int mb_find_order_for_block(struct ext4_buddy *e4b, int block) | |
1249 | { | |
1250 | int order = 1; | |
1251 | void *bb; | |
1252 | ||
c5e8f3f3 | 1253 | BUG_ON(e4b->bd_bitmap == e4b->bd_buddy); |
c9de560d AT |
1254 | BUG_ON(block >= (1 << (e4b->bd_blkbits + 3))); |
1255 | ||
c5e8f3f3 | 1256 | bb = e4b->bd_buddy; |
c9de560d AT |
1257 | while (order <= e4b->bd_blkbits + 1) { |
1258 | block = block >> 1; | |
1259 | if (!mb_test_bit(block, bb)) { | |
1260 | /* this block is part of buddy of order 'order' */ | |
1261 | return order; | |
1262 | } | |
1263 | bb += 1 << (e4b->bd_blkbits - order); | |
1264 | order++; | |
1265 | } | |
1266 | return 0; | |
1267 | } | |
1268 | ||
955ce5f5 | 1269 | static void mb_clear_bits(void *bm, int cur, int len) |
c9de560d AT |
1270 | { |
1271 | __u32 *addr; | |
1272 | ||
1273 | len = cur + len; | |
1274 | while (cur < len) { | |
1275 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1276 | /* fast path: clear whole word at once */ | |
1277 | addr = bm + (cur >> 3); | |
1278 | *addr = 0; | |
1279 | cur += 32; | |
1280 | continue; | |
1281 | } | |
955ce5f5 | 1282 | mb_clear_bit(cur, bm); |
c9de560d AT |
1283 | cur++; |
1284 | } | |
1285 | } | |
1286 | ||
eabe0444 AS |
1287 | /* clear bits in given range |
1288 | * will return first found zero bit if any, -1 otherwise | |
1289 | */ | |
1290 | static int mb_test_and_clear_bits(void *bm, int cur, int len) | |
1291 | { | |
1292 | __u32 *addr; | |
1293 | int zero_bit = -1; | |
1294 | ||
1295 | len = cur + len; | |
1296 | while (cur < len) { | |
1297 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1298 | /* fast path: clear whole word at once */ | |
1299 | addr = bm + (cur >> 3); | |
1300 | if (*addr != (__u32)(-1) && zero_bit == -1) | |
1301 | zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0); | |
1302 | *addr = 0; | |
1303 | cur += 32; | |
1304 | continue; | |
1305 | } | |
1306 | if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1) | |
1307 | zero_bit = cur; | |
1308 | cur++; | |
1309 | } | |
1310 | ||
1311 | return zero_bit; | |
1312 | } | |
1313 | ||
c3e94d1d | 1314 | void ext4_set_bits(void *bm, int cur, int len) |
c9de560d AT |
1315 | { |
1316 | __u32 *addr; | |
1317 | ||
1318 | len = cur + len; | |
1319 | while (cur < len) { | |
1320 | if ((cur & 31) == 0 && (len - cur) >= 32) { | |
1321 | /* fast path: set whole word at once */ | |
1322 | addr = bm + (cur >> 3); | |
1323 | *addr = 0xffffffff; | |
1324 | cur += 32; | |
1325 | continue; | |
1326 | } | |
955ce5f5 | 1327 | mb_set_bit(cur, bm); |
c9de560d AT |
1328 | cur++; |
1329 | } | |
1330 | } | |
1331 | ||
eabe0444 AS |
1332 | /* |
1333 | * _________________________________________________________________ */ | |
1334 | ||
1335 | static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side) | |
1336 | { | |
1337 | if (mb_test_bit(*bit + side, bitmap)) { | |
1338 | mb_clear_bit(*bit, bitmap); | |
1339 | (*bit) -= side; | |
1340 | return 1; | |
1341 | } | |
1342 | else { | |
1343 | (*bit) += side; | |
1344 | mb_set_bit(*bit, bitmap); | |
1345 | return -1; | |
1346 | } | |
1347 | } | |
1348 | ||
1349 | static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last) | |
1350 | { | |
1351 | int max; | |
1352 | int order = 1; | |
1353 | void *buddy = mb_find_buddy(e4b, order, &max); | |
1354 | ||
1355 | while (buddy) { | |
1356 | void *buddy2; | |
1357 | ||
1358 | /* Bits in range [first; last] are known to be set since | |
1359 | * corresponding blocks were allocated. Bits in range | |
1360 | * (first; last) will stay set because they form buddies on | |
1361 | * upper layer. We just deal with borders if they don't | |
1362 | * align with upper layer and then go up. | |
1363 | * Releasing entire group is all about clearing | |
1364 | * single bit of highest order buddy. | |
1365 | */ | |
1366 | ||
1367 | /* Example: | |
1368 | * --------------------------------- | |
1369 | * | 1 | 1 | 1 | 1 | | |
1370 | * --------------------------------- | |
1371 | * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | | |
1372 | * --------------------------------- | |
1373 | * 0 1 2 3 4 5 6 7 | |
1374 | * \_____________________/ | |
1375 | * | |
1376 | * Neither [1] nor [6] is aligned to above layer. | |
1377 | * Left neighbour [0] is free, so mark it busy, | |
1378 | * decrease bb_counters and extend range to | |
1379 | * [0; 6] | |
1380 | * Right neighbour [7] is busy. It can't be coaleasced with [6], so | |
1381 | * mark [6] free, increase bb_counters and shrink range to | |
1382 | * [0; 5]. | |
1383 | * Then shift range to [0; 2], go up and do the same. | |
1384 | */ | |
1385 | ||
1386 | ||
1387 | if (first & 1) | |
1388 | e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1); | |
1389 | if (!(last & 1)) | |
1390 | e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1); | |
1391 | if (first > last) | |
1392 | break; | |
1393 | order++; | |
1394 | ||
1395 | if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) { | |
1396 | mb_clear_bits(buddy, first, last - first + 1); | |
1397 | e4b->bd_info->bb_counters[order - 1] += last - first + 1; | |
1398 | break; | |
1399 | } | |
1400 | first >>= 1; | |
1401 | last >>= 1; | |
1402 | buddy = buddy2; | |
1403 | } | |
1404 | } | |
1405 | ||
7e5a8cdd | 1406 | static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b, |
eabe0444 | 1407 | int first, int count) |
c9de560d | 1408 | { |
eabe0444 AS |
1409 | int left_is_free = 0; |
1410 | int right_is_free = 0; | |
1411 | int block; | |
1412 | int last = first + count - 1; | |
c9de560d AT |
1413 | struct super_block *sb = e4b->bd_sb; |
1414 | ||
eabe0444 | 1415 | BUG_ON(last >= (sb->s_blocksize << 3)); |
bc8e6740 | 1416 | assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); |
163a203d DW |
1417 | /* Don't bother if the block group is corrupt. */ |
1418 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) | |
1419 | return; | |
1420 | ||
c9de560d AT |
1421 | mb_check_buddy(e4b); |
1422 | mb_free_blocks_double(inode, e4b, first, count); | |
1423 | ||
1424 | e4b->bd_info->bb_free += count; | |
1425 | if (first < e4b->bd_info->bb_first_free) | |
1426 | e4b->bd_info->bb_first_free = first; | |
1427 | ||
eabe0444 AS |
1428 | /* access memory sequentially: check left neighbour, |
1429 | * clear range and then check right neighbour | |
1430 | */ | |
c9de560d | 1431 | if (first != 0) |
eabe0444 AS |
1432 | left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap); |
1433 | block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count); | |
1434 | if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0]) | |
1435 | right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap); | |
1436 | ||
1437 | if (unlikely(block != -1)) { | |
e43bb4e6 | 1438 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
eabe0444 AS |
1439 | ext4_fsblk_t blocknr; |
1440 | ||
1441 | blocknr = ext4_group_first_block_no(sb, e4b->bd_group); | |
1442 | blocknr += EXT4_C2B(EXT4_SB(sb), block); | |
1443 | ext4_grp_locked_error(sb, e4b->bd_group, | |
1444 | inode ? inode->i_ino : 0, | |
1445 | blocknr, | |
1446 | "freeing already freed block " | |
163a203d DW |
1447 | "(bit %u); block bitmap corrupt.", |
1448 | block); | |
e43bb4e6 NJ |
1449 | if (!EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)) |
1450 | percpu_counter_sub(&sbi->s_freeclusters_counter, | |
1451 | e4b->bd_info->bb_free); | |
163a203d DW |
1452 | /* Mark the block group as corrupt. */ |
1453 | set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT, | |
1454 | &e4b->bd_info->bb_state); | |
eabe0444 AS |
1455 | mb_regenerate_buddy(e4b); |
1456 | goto done; | |
1457 | } | |
1458 | ||
1459 | /* let's maintain fragments counter */ | |
1460 | if (left_is_free && right_is_free) | |
c9de560d | 1461 | e4b->bd_info->bb_fragments--; |
eabe0444 | 1462 | else if (!left_is_free && !right_is_free) |
c9de560d AT |
1463 | e4b->bd_info->bb_fragments++; |
1464 | ||
eabe0444 AS |
1465 | /* buddy[0] == bd_bitmap is a special case, so handle |
1466 | * it right away and let mb_buddy_mark_free stay free of | |
1467 | * zero order checks. | |
1468 | * Check if neighbours are to be coaleasced, | |
1469 | * adjust bitmap bb_counters and borders appropriately. | |
1470 | */ | |
1471 | if (first & 1) { | |
1472 | first += !left_is_free; | |
1473 | e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1; | |
1474 | } | |
1475 | if (!(last & 1)) { | |
1476 | last -= !right_is_free; | |
1477 | e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1; | |
1478 | } | |
c9de560d | 1479 | |
eabe0444 AS |
1480 | if (first <= last) |
1481 | mb_buddy_mark_free(e4b, first >> 1, last >> 1); | |
c9de560d | 1482 | |
eabe0444 | 1483 | done: |
8a57d9d6 | 1484 | mb_set_largest_free_order(sb, e4b->bd_info); |
c9de560d | 1485 | mb_check_buddy(e4b); |
c9de560d AT |
1486 | } |
1487 | ||
15c006a2 | 1488 | static int mb_find_extent(struct ext4_buddy *e4b, int block, |
c9de560d AT |
1489 | int needed, struct ext4_free_extent *ex) |
1490 | { | |
1491 | int next = block; | |
15c006a2 | 1492 | int max, order; |
c9de560d AT |
1493 | void *buddy; |
1494 | ||
bc8e6740 | 1495 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
1496 | BUG_ON(ex == NULL); |
1497 | ||
15c006a2 | 1498 | buddy = mb_find_buddy(e4b, 0, &max); |
c9de560d AT |
1499 | BUG_ON(buddy == NULL); |
1500 | BUG_ON(block >= max); | |
1501 | if (mb_test_bit(block, buddy)) { | |
1502 | ex->fe_len = 0; | |
1503 | ex->fe_start = 0; | |
1504 | ex->fe_group = 0; | |
1505 | return 0; | |
1506 | } | |
1507 | ||
15c006a2 RD |
1508 | /* find actual order */ |
1509 | order = mb_find_order_for_block(e4b, block); | |
1510 | block = block >> order; | |
c9de560d AT |
1511 | |
1512 | ex->fe_len = 1 << order; | |
1513 | ex->fe_start = block << order; | |
1514 | ex->fe_group = e4b->bd_group; | |
1515 | ||
1516 | /* calc difference from given start */ | |
1517 | next = next - ex->fe_start; | |
1518 | ex->fe_len -= next; | |
1519 | ex->fe_start += next; | |
1520 | ||
1521 | while (needed > ex->fe_len && | |
d8ec0c39 | 1522 | mb_find_buddy(e4b, order, &max)) { |
c9de560d AT |
1523 | |
1524 | if (block + 1 >= max) | |
1525 | break; | |
1526 | ||
1527 | next = (block + 1) * (1 << order); | |
c5e8f3f3 | 1528 | if (mb_test_bit(next, e4b->bd_bitmap)) |
c9de560d AT |
1529 | break; |
1530 | ||
b051d8dc | 1531 | order = mb_find_order_for_block(e4b, next); |
c9de560d | 1532 | |
c9de560d AT |
1533 | block = next >> order; |
1534 | ex->fe_len += 1 << order; | |
1535 | } | |
1536 | ||
1537 | BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3))); | |
1538 | return ex->fe_len; | |
1539 | } | |
1540 | ||
1541 | static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex) | |
1542 | { | |
1543 | int ord; | |
1544 | int mlen = 0; | |
1545 | int max = 0; | |
1546 | int cur; | |
1547 | int start = ex->fe_start; | |
1548 | int len = ex->fe_len; | |
1549 | unsigned ret = 0; | |
1550 | int len0 = len; | |
1551 | void *buddy; | |
1552 | ||
1553 | BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3)); | |
1554 | BUG_ON(e4b->bd_group != ex->fe_group); | |
bc8e6740 | 1555 | assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); |
c9de560d AT |
1556 | mb_check_buddy(e4b); |
1557 | mb_mark_used_double(e4b, start, len); | |
1558 | ||
1559 | e4b->bd_info->bb_free -= len; | |
1560 | if (e4b->bd_info->bb_first_free == start) | |
1561 | e4b->bd_info->bb_first_free += len; | |
1562 | ||
1563 | /* let's maintain fragments counter */ | |
1564 | if (start != 0) | |
c5e8f3f3 | 1565 | mlen = !mb_test_bit(start - 1, e4b->bd_bitmap); |
c9de560d | 1566 | if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0]) |
c5e8f3f3 | 1567 | max = !mb_test_bit(start + len, e4b->bd_bitmap); |
c9de560d AT |
1568 | if (mlen && max) |
1569 | e4b->bd_info->bb_fragments++; | |
1570 | else if (!mlen && !max) | |
1571 | e4b->bd_info->bb_fragments--; | |
1572 | ||
1573 | /* let's maintain buddy itself */ | |
1574 | while (len) { | |
1575 | ord = mb_find_order_for_block(e4b, start); | |
1576 | ||
1577 | if (((start >> ord) << ord) == start && len >= (1 << ord)) { | |
1578 | /* the whole chunk may be allocated at once! */ | |
1579 | mlen = 1 << ord; | |
1580 | buddy = mb_find_buddy(e4b, ord, &max); | |
1581 | BUG_ON((start >> ord) >= max); | |
1582 | mb_set_bit(start >> ord, buddy); | |
1583 | e4b->bd_info->bb_counters[ord]--; | |
1584 | start += mlen; | |
1585 | len -= mlen; | |
1586 | BUG_ON(len < 0); | |
1587 | continue; | |
1588 | } | |
1589 | ||
1590 | /* store for history */ | |
1591 | if (ret == 0) | |
1592 | ret = len | (ord << 16); | |
1593 | ||
1594 | /* we have to split large buddy */ | |
1595 | BUG_ON(ord <= 0); | |
1596 | buddy = mb_find_buddy(e4b, ord, &max); | |
1597 | mb_set_bit(start >> ord, buddy); | |
1598 | e4b->bd_info->bb_counters[ord]--; | |
1599 | ||
1600 | ord--; | |
1601 | cur = (start >> ord) & ~1U; | |
1602 | buddy = mb_find_buddy(e4b, ord, &max); | |
1603 | mb_clear_bit(cur, buddy); | |
1604 | mb_clear_bit(cur + 1, buddy); | |
1605 | e4b->bd_info->bb_counters[ord]++; | |
1606 | e4b->bd_info->bb_counters[ord]++; | |
1607 | } | |
8a57d9d6 | 1608 | mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info); |
c9de560d | 1609 | |
c5e8f3f3 | 1610 | ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0); |
c9de560d AT |
1611 | mb_check_buddy(e4b); |
1612 | ||
1613 | return ret; | |
1614 | } | |
1615 | ||
1616 | /* | |
1617 | * Must be called under group lock! | |
1618 | */ | |
1619 | static void ext4_mb_use_best_found(struct ext4_allocation_context *ac, | |
1620 | struct ext4_buddy *e4b) | |
1621 | { | |
1622 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
1623 | int ret; | |
1624 | ||
1625 | BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group); | |
1626 | BUG_ON(ac->ac_status == AC_STATUS_FOUND); | |
1627 | ||
1628 | ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len); | |
1629 | ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical; | |
1630 | ret = mb_mark_used(e4b, &ac->ac_b_ex); | |
1631 | ||
1632 | /* preallocation can change ac_b_ex, thus we store actually | |
1633 | * allocated blocks for history */ | |
1634 | ac->ac_f_ex = ac->ac_b_ex; | |
1635 | ||
1636 | ac->ac_status = AC_STATUS_FOUND; | |
1637 | ac->ac_tail = ret & 0xffff; | |
1638 | ac->ac_buddy = ret >> 16; | |
1639 | ||
c3a326a6 AK |
1640 | /* |
1641 | * take the page reference. We want the page to be pinned | |
1642 | * so that we don't get a ext4_mb_init_cache_call for this | |
1643 | * group until we update the bitmap. That would mean we | |
1644 | * double allocate blocks. The reference is dropped | |
1645 | * in ext4_mb_release_context | |
1646 | */ | |
c9de560d AT |
1647 | ac->ac_bitmap_page = e4b->bd_bitmap_page; |
1648 | get_page(ac->ac_bitmap_page); | |
1649 | ac->ac_buddy_page = e4b->bd_buddy_page; | |
1650 | get_page(ac->ac_buddy_page); | |
c9de560d | 1651 | /* store last allocated for subsequent stream allocation */ |
4ba74d00 | 1652 | if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { |
c9de560d AT |
1653 | spin_lock(&sbi->s_md_lock); |
1654 | sbi->s_mb_last_group = ac->ac_f_ex.fe_group; | |
1655 | sbi->s_mb_last_start = ac->ac_f_ex.fe_start; | |
1656 | spin_unlock(&sbi->s_md_lock); | |
1657 | } | |
1658 | } | |
1659 | ||
1660 | /* | |
1661 | * regular allocator, for general purposes allocation | |
1662 | */ | |
1663 | ||
1664 | static void ext4_mb_check_limits(struct ext4_allocation_context *ac, | |
1665 | struct ext4_buddy *e4b, | |
1666 | int finish_group) | |
1667 | { | |
1668 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
1669 | struct ext4_free_extent *bex = &ac->ac_b_ex; | |
1670 | struct ext4_free_extent *gex = &ac->ac_g_ex; | |
1671 | struct ext4_free_extent ex; | |
1672 | int max; | |
1673 | ||
032115fc AK |
1674 | if (ac->ac_status == AC_STATUS_FOUND) |
1675 | return; | |
c9de560d AT |
1676 | /* |
1677 | * We don't want to scan for a whole year | |
1678 | */ | |
1679 | if (ac->ac_found > sbi->s_mb_max_to_scan && | |
1680 | !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
1681 | ac->ac_status = AC_STATUS_BREAK; | |
1682 | return; | |
1683 | } | |
1684 | ||
1685 | /* | |
1686 | * Haven't found good chunk so far, let's continue | |
1687 | */ | |
1688 | if (bex->fe_len < gex->fe_len) | |
1689 | return; | |
1690 | ||
1691 | if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan) | |
1692 | && bex->fe_group == e4b->bd_group) { | |
1693 | /* recheck chunk's availability - we don't know | |
1694 | * when it was found (within this lock-unlock | |
1695 | * period or not) */ | |
15c006a2 | 1696 | max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex); |
c9de560d AT |
1697 | if (max >= gex->fe_len) { |
1698 | ext4_mb_use_best_found(ac, e4b); | |
1699 | return; | |
1700 | } | |
1701 | } | |
1702 | } | |
1703 | ||
1704 | /* | |
1705 | * The routine checks whether found extent is good enough. If it is, | |
1706 | * then the extent gets marked used and flag is set to the context | |
1707 | * to stop scanning. Otherwise, the extent is compared with the | |
1708 | * previous found extent and if new one is better, then it's stored | |
1709 | * in the context. Later, the best found extent will be used, if | |
1710 | * mballoc can't find good enough extent. | |
1711 | * | |
1712 | * FIXME: real allocation policy is to be designed yet! | |
1713 | */ | |
1714 | static void ext4_mb_measure_extent(struct ext4_allocation_context *ac, | |
1715 | struct ext4_free_extent *ex, | |
1716 | struct ext4_buddy *e4b) | |
1717 | { | |
1718 | struct ext4_free_extent *bex = &ac->ac_b_ex; | |
1719 | struct ext4_free_extent *gex = &ac->ac_g_ex; | |
1720 | ||
1721 | BUG_ON(ex->fe_len <= 0); | |
7137d7a4 TT |
1722 | BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); |
1723 | BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); | |
c9de560d AT |
1724 | BUG_ON(ac->ac_status != AC_STATUS_CONTINUE); |
1725 | ||
1726 | ac->ac_found++; | |
1727 | ||
1728 | /* | |
1729 | * The special case - take what you catch first | |
1730 | */ | |
1731 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
1732 | *bex = *ex; | |
1733 | ext4_mb_use_best_found(ac, e4b); | |
1734 | return; | |
1735 | } | |
1736 | ||
1737 | /* | |
1738 | * Let's check whether the chuck is good enough | |
1739 | */ | |
1740 | if (ex->fe_len == gex->fe_len) { | |
1741 | *bex = *ex; | |
1742 | ext4_mb_use_best_found(ac, e4b); | |
1743 | return; | |
1744 | } | |
1745 | ||
1746 | /* | |
1747 | * If this is first found extent, just store it in the context | |
1748 | */ | |
1749 | if (bex->fe_len == 0) { | |
1750 | *bex = *ex; | |
1751 | return; | |
1752 | } | |
1753 | ||
1754 | /* | |
1755 | * If new found extent is better, store it in the context | |
1756 | */ | |
1757 | if (bex->fe_len < gex->fe_len) { | |
1758 | /* if the request isn't satisfied, any found extent | |
1759 | * larger than previous best one is better */ | |
1760 | if (ex->fe_len > bex->fe_len) | |
1761 | *bex = *ex; | |
1762 | } else if (ex->fe_len > gex->fe_len) { | |
1763 | /* if the request is satisfied, then we try to find | |
1764 | * an extent that still satisfy the request, but is | |
1765 | * smaller than previous one */ | |
1766 | if (ex->fe_len < bex->fe_len) | |
1767 | *bex = *ex; | |
1768 | } | |
1769 | ||
1770 | ext4_mb_check_limits(ac, e4b, 0); | |
1771 | } | |
1772 | ||
089ceecc ES |
1773 | static noinline_for_stack |
1774 | int ext4_mb_try_best_found(struct ext4_allocation_context *ac, | |
c9de560d AT |
1775 | struct ext4_buddy *e4b) |
1776 | { | |
1777 | struct ext4_free_extent ex = ac->ac_b_ex; | |
1778 | ext4_group_t group = ex.fe_group; | |
1779 | int max; | |
1780 | int err; | |
1781 | ||
1782 | BUG_ON(ex.fe_len <= 0); | |
1783 | err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); | |
1784 | if (err) | |
1785 | return err; | |
1786 | ||
1787 | ext4_lock_group(ac->ac_sb, group); | |
15c006a2 | 1788 | max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex); |
c9de560d AT |
1789 | |
1790 | if (max > 0) { | |
1791 | ac->ac_b_ex = ex; | |
1792 | ext4_mb_use_best_found(ac, e4b); | |
1793 | } | |
1794 | ||
1795 | ext4_unlock_group(ac->ac_sb, group); | |
e39e07fd | 1796 | ext4_mb_unload_buddy(e4b); |
c9de560d AT |
1797 | |
1798 | return 0; | |
1799 | } | |
1800 | ||
089ceecc ES |
1801 | static noinline_for_stack |
1802 | int ext4_mb_find_by_goal(struct ext4_allocation_context *ac, | |
c9de560d AT |
1803 | struct ext4_buddy *e4b) |
1804 | { | |
1805 | ext4_group_t group = ac->ac_g_ex.fe_group; | |
1806 | int max; | |
1807 | int err; | |
1808 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
838cd0cf | 1809 | struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); |
c9de560d AT |
1810 | struct ext4_free_extent ex; |
1811 | ||
1812 | if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL)) | |
1813 | return 0; | |
838cd0cf YY |
1814 | if (grp->bb_free == 0) |
1815 | return 0; | |
c9de560d AT |
1816 | |
1817 | err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); | |
1818 | if (err) | |
1819 | return err; | |
1820 | ||
163a203d DW |
1821 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) { |
1822 | ext4_mb_unload_buddy(e4b); | |
1823 | return 0; | |
1824 | } | |
1825 | ||
c9de560d | 1826 | ext4_lock_group(ac->ac_sb, group); |
15c006a2 | 1827 | max = mb_find_extent(e4b, ac->ac_g_ex.fe_start, |
c9de560d | 1828 | ac->ac_g_ex.fe_len, &ex); |
ab0c00fc | 1829 | ex.fe_logical = 0xDEADFA11; /* debug value */ |
c9de560d AT |
1830 | |
1831 | if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) { | |
1832 | ext4_fsblk_t start; | |
1833 | ||
5661bd68 AM |
1834 | start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) + |
1835 | ex.fe_start; | |
c9de560d AT |
1836 | /* use do_div to get remainder (would be 64-bit modulo) */ |
1837 | if (do_div(start, sbi->s_stripe) == 0) { | |
1838 | ac->ac_found++; | |
1839 | ac->ac_b_ex = ex; | |
1840 | ext4_mb_use_best_found(ac, e4b); | |
1841 | } | |
1842 | } else if (max >= ac->ac_g_ex.fe_len) { | |
1843 | BUG_ON(ex.fe_len <= 0); | |
1844 | BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); | |
1845 | BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); | |
1846 | ac->ac_found++; | |
1847 | ac->ac_b_ex = ex; | |
1848 | ext4_mb_use_best_found(ac, e4b); | |
1849 | } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) { | |
1850 | /* Sometimes, caller may want to merge even small | |
1851 | * number of blocks to an existing extent */ | |
1852 | BUG_ON(ex.fe_len <= 0); | |
1853 | BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); | |
1854 | BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); | |
1855 | ac->ac_found++; | |
1856 | ac->ac_b_ex = ex; | |
1857 | ext4_mb_use_best_found(ac, e4b); | |
1858 | } | |
1859 | ext4_unlock_group(ac->ac_sb, group); | |
e39e07fd | 1860 | ext4_mb_unload_buddy(e4b); |
c9de560d AT |
1861 | |
1862 | return 0; | |
1863 | } | |
1864 | ||
1865 | /* | |
1866 | * The routine scans buddy structures (not bitmap!) from given order | |
1867 | * to max order and tries to find big enough chunk to satisfy the req | |
1868 | */ | |
089ceecc ES |
1869 | static noinline_for_stack |
1870 | void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac, | |
c9de560d AT |
1871 | struct ext4_buddy *e4b) |
1872 | { | |
1873 | struct super_block *sb = ac->ac_sb; | |
1874 | struct ext4_group_info *grp = e4b->bd_info; | |
1875 | void *buddy; | |
1876 | int i; | |
1877 | int k; | |
1878 | int max; | |
1879 | ||
1880 | BUG_ON(ac->ac_2order <= 0); | |
1881 | for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) { | |
1882 | if (grp->bb_counters[i] == 0) | |
1883 | continue; | |
1884 | ||
1885 | buddy = mb_find_buddy(e4b, i, &max); | |
1886 | BUG_ON(buddy == NULL); | |
1887 | ||
ffad0a44 | 1888 | k = mb_find_next_zero_bit(buddy, max, 0); |
c9de560d AT |
1889 | BUG_ON(k >= max); |
1890 | ||
1891 | ac->ac_found++; | |
1892 | ||
1893 | ac->ac_b_ex.fe_len = 1 << i; | |
1894 | ac->ac_b_ex.fe_start = k << i; | |
1895 | ac->ac_b_ex.fe_group = e4b->bd_group; | |
1896 | ||
1897 | ext4_mb_use_best_found(ac, e4b); | |
1898 | ||
1899 | BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len); | |
1900 | ||
1901 | if (EXT4_SB(sb)->s_mb_stats) | |
1902 | atomic_inc(&EXT4_SB(sb)->s_bal_2orders); | |
1903 | ||
1904 | break; | |
1905 | } | |
1906 | } | |
1907 | ||
1908 | /* | |
1909 | * The routine scans the group and measures all found extents. | |
1910 | * In order to optimize scanning, caller must pass number of | |
1911 | * free blocks in the group, so the routine can know upper limit. | |
1912 | */ | |
089ceecc ES |
1913 | static noinline_for_stack |
1914 | void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac, | |
c9de560d AT |
1915 | struct ext4_buddy *e4b) |
1916 | { | |
1917 | struct super_block *sb = ac->ac_sb; | |
c5e8f3f3 | 1918 | void *bitmap = e4b->bd_bitmap; |
c9de560d AT |
1919 | struct ext4_free_extent ex; |
1920 | int i; | |
1921 | int free; | |
1922 | ||
1923 | free = e4b->bd_info->bb_free; | |
1924 | BUG_ON(free <= 0); | |
1925 | ||
1926 | i = e4b->bd_info->bb_first_free; | |
1927 | ||
1928 | while (free && ac->ac_status == AC_STATUS_CONTINUE) { | |
ffad0a44 | 1929 | i = mb_find_next_zero_bit(bitmap, |
7137d7a4 TT |
1930 | EXT4_CLUSTERS_PER_GROUP(sb), i); |
1931 | if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) { | |
26346ff6 | 1932 | /* |
e56eb659 | 1933 | * IF we have corrupt bitmap, we won't find any |
26346ff6 AK |
1934 | * free blocks even though group info says we |
1935 | * we have free blocks | |
1936 | */ | |
e29136f8 | 1937 | ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, |
53accfa9 | 1938 | "%d free clusters as per " |
fde4d95a | 1939 | "group info. But bitmap says 0", |
26346ff6 | 1940 | free); |
c9de560d AT |
1941 | break; |
1942 | } | |
1943 | ||
15c006a2 | 1944 | mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex); |
c9de560d | 1945 | BUG_ON(ex.fe_len <= 0); |
26346ff6 | 1946 | if (free < ex.fe_len) { |
e29136f8 | 1947 | ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, |
53accfa9 | 1948 | "%d free clusters as per " |
fde4d95a | 1949 | "group info. But got %d blocks", |
26346ff6 | 1950 | free, ex.fe_len); |
e56eb659 AK |
1951 | /* |
1952 | * The number of free blocks differs. This mostly | |
1953 | * indicate that the bitmap is corrupt. So exit | |
1954 | * without claiming the space. | |
1955 | */ | |
1956 | break; | |
26346ff6 | 1957 | } |
ab0c00fc | 1958 | ex.fe_logical = 0xDEADC0DE; /* debug value */ |
c9de560d AT |
1959 | ext4_mb_measure_extent(ac, &ex, e4b); |
1960 | ||
1961 | i += ex.fe_len; | |
1962 | free -= ex.fe_len; | |
1963 | } | |
1964 | ||
1965 | ext4_mb_check_limits(ac, e4b, 1); | |
1966 | } | |
1967 | ||
1968 | /* | |
1969 | * This is a special case for storages like raid5 | |
506bf2d8 | 1970 | * we try to find stripe-aligned chunks for stripe-size-multiple requests |
c9de560d | 1971 | */ |
089ceecc ES |
1972 | static noinline_for_stack |
1973 | void ext4_mb_scan_aligned(struct ext4_allocation_context *ac, | |
c9de560d AT |
1974 | struct ext4_buddy *e4b) |
1975 | { | |
1976 | struct super_block *sb = ac->ac_sb; | |
1977 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
c5e8f3f3 | 1978 | void *bitmap = e4b->bd_bitmap; |
c9de560d AT |
1979 | struct ext4_free_extent ex; |
1980 | ext4_fsblk_t first_group_block; | |
1981 | ext4_fsblk_t a; | |
1982 | ext4_grpblk_t i; | |
1983 | int max; | |
1984 | ||
1985 | BUG_ON(sbi->s_stripe == 0); | |
1986 | ||
1987 | /* find first stripe-aligned block in group */ | |
5661bd68 AM |
1988 | first_group_block = ext4_group_first_block_no(sb, e4b->bd_group); |
1989 | ||
c9de560d AT |
1990 | a = first_group_block + sbi->s_stripe - 1; |
1991 | do_div(a, sbi->s_stripe); | |
1992 | i = (a * sbi->s_stripe) - first_group_block; | |
1993 | ||
7137d7a4 | 1994 | while (i < EXT4_CLUSTERS_PER_GROUP(sb)) { |
c9de560d | 1995 | if (!mb_test_bit(i, bitmap)) { |
15c006a2 | 1996 | max = mb_find_extent(e4b, i, sbi->s_stripe, &ex); |
c9de560d AT |
1997 | if (max >= sbi->s_stripe) { |
1998 | ac->ac_found++; | |
ab0c00fc | 1999 | ex.fe_logical = 0xDEADF00D; /* debug value */ |
c9de560d AT |
2000 | ac->ac_b_ex = ex; |
2001 | ext4_mb_use_best_found(ac, e4b); | |
2002 | break; | |
2003 | } | |
2004 | } | |
2005 | i += sbi->s_stripe; | |
2006 | } | |
2007 | } | |
2008 | ||
8a57d9d6 | 2009 | /* This is now called BEFORE we load the buddy bitmap. */ |
c9de560d AT |
2010 | static int ext4_mb_good_group(struct ext4_allocation_context *ac, |
2011 | ext4_group_t group, int cr) | |
2012 | { | |
2013 | unsigned free, fragments; | |
a4912123 | 2014 | int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb)); |
c9de560d AT |
2015 | struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); |
2016 | ||
2017 | BUG_ON(cr < 0 || cr >= 4); | |
8a57d9d6 | 2018 | |
01fc48e8 TT |
2019 | free = grp->bb_free; |
2020 | if (free == 0) | |
2021 | return 0; | |
2022 | if (cr <= 2 && free < ac->ac_g_ex.fe_len) | |
2023 | return 0; | |
2024 | ||
163a203d DW |
2025 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp))) |
2026 | return 0; | |
2027 | ||
8a57d9d6 CW |
2028 | /* We only do this if the grp has never been initialized */ |
2029 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { | |
2030 | int ret = ext4_mb_init_group(ac->ac_sb, group); | |
2031 | if (ret) | |
2032 | return 0; | |
2033 | } | |
c9de560d | 2034 | |
c9de560d | 2035 | fragments = grp->bb_fragments; |
c9de560d AT |
2036 | if (fragments == 0) |
2037 | return 0; | |
2038 | ||
2039 | switch (cr) { | |
2040 | case 0: | |
2041 | BUG_ON(ac->ac_2order == 0); | |
c9de560d | 2042 | |
a4912123 TT |
2043 | /* Avoid using the first bg of a flexgroup for data files */ |
2044 | if ((ac->ac_flags & EXT4_MB_HINT_DATA) && | |
2045 | (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) && | |
2046 | ((group % flex_size) == 0)) | |
2047 | return 0; | |
2048 | ||
40ae3487 TT |
2049 | if ((ac->ac_2order > ac->ac_sb->s_blocksize_bits+1) || |
2050 | (free / fragments) >= ac->ac_g_ex.fe_len) | |
2051 | return 1; | |
2052 | ||
2053 | if (grp->bb_largest_free_order < ac->ac_2order) | |
2054 | return 0; | |
2055 | ||
8a57d9d6 | 2056 | return 1; |
c9de560d AT |
2057 | case 1: |
2058 | if ((free / fragments) >= ac->ac_g_ex.fe_len) | |
2059 | return 1; | |
2060 | break; | |
2061 | case 2: | |
2062 | if (free >= ac->ac_g_ex.fe_len) | |
2063 | return 1; | |
2064 | break; | |
2065 | case 3: | |
2066 | return 1; | |
2067 | default: | |
2068 | BUG(); | |
2069 | } | |
2070 | ||
2071 | return 0; | |
2072 | } | |
2073 | ||
4ddfef7b ES |
2074 | static noinline_for_stack int |
2075 | ext4_mb_regular_allocator(struct ext4_allocation_context *ac) | |
c9de560d | 2076 | { |
8df9675f | 2077 | ext4_group_t ngroups, group, i; |
c9de560d AT |
2078 | int cr; |
2079 | int err = 0; | |
c9de560d AT |
2080 | struct ext4_sb_info *sbi; |
2081 | struct super_block *sb; | |
2082 | struct ext4_buddy e4b; | |
c9de560d AT |
2083 | |
2084 | sb = ac->ac_sb; | |
2085 | sbi = EXT4_SB(sb); | |
8df9675f | 2086 | ngroups = ext4_get_groups_count(sb); |
fb0a387d | 2087 | /* non-extent files are limited to low blocks/groups */ |
12e9b892 | 2088 | if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS))) |
fb0a387d ES |
2089 | ngroups = sbi->s_blockfile_groups; |
2090 | ||
c9de560d AT |
2091 | BUG_ON(ac->ac_status == AC_STATUS_FOUND); |
2092 | ||
2093 | /* first, try the goal */ | |
2094 | err = ext4_mb_find_by_goal(ac, &e4b); | |
2095 | if (err || ac->ac_status == AC_STATUS_FOUND) | |
2096 | goto out; | |
2097 | ||
2098 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) | |
2099 | goto out; | |
2100 | ||
2101 | /* | |
2102 | * ac->ac2_order is set only if the fe_len is a power of 2 | |
2103 | * if ac2_order is set we also set criteria to 0 so that we | |
2104 | * try exact allocation using buddy. | |
2105 | */ | |
2106 | i = fls(ac->ac_g_ex.fe_len); | |
2107 | ac->ac_2order = 0; | |
2108 | /* | |
2109 | * We search using buddy data only if the order of the request | |
2110 | * is greater than equal to the sbi_s_mb_order2_reqs | |
b713a5ec | 2111 | * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req |
c9de560d AT |
2112 | */ |
2113 | if (i >= sbi->s_mb_order2_reqs) { | |
2114 | /* | |
2115 | * This should tell if fe_len is exactly power of 2 | |
2116 | */ | |
2117 | if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0) | |
2118 | ac->ac_2order = i - 1; | |
2119 | } | |
2120 | ||
4ba74d00 TT |
2121 | /* if stream allocation is enabled, use global goal */ |
2122 | if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { | |
c9de560d AT |
2123 | /* TBD: may be hot point */ |
2124 | spin_lock(&sbi->s_md_lock); | |
2125 | ac->ac_g_ex.fe_group = sbi->s_mb_last_group; | |
2126 | ac->ac_g_ex.fe_start = sbi->s_mb_last_start; | |
2127 | spin_unlock(&sbi->s_md_lock); | |
2128 | } | |
4ba74d00 | 2129 | |
c9de560d AT |
2130 | /* Let's just scan groups to find more-less suitable blocks */ |
2131 | cr = ac->ac_2order ? 0 : 1; | |
2132 | /* | |
2133 | * cr == 0 try to get exact allocation, | |
2134 | * cr == 3 try to get anything | |
2135 | */ | |
2136 | repeat: | |
2137 | for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) { | |
2138 | ac->ac_criteria = cr; | |
ed8f9c75 AK |
2139 | /* |
2140 | * searching for the right group start | |
2141 | * from the goal value specified | |
2142 | */ | |
2143 | group = ac->ac_g_ex.fe_group; | |
2144 | ||
8df9675f | 2145 | for (i = 0; i < ngroups; group++, i++) { |
2ed5724d | 2146 | cond_resched(); |
e6155736 LM |
2147 | /* |
2148 | * Artificially restricted ngroups for non-extent | |
2149 | * files makes group > ngroups possible on first loop. | |
2150 | */ | |
2151 | if (group >= ngroups) | |
c9de560d AT |
2152 | group = 0; |
2153 | ||
8a57d9d6 CW |
2154 | /* This now checks without needing the buddy page */ |
2155 | if (!ext4_mb_good_group(ac, group, cr)) | |
c9de560d AT |
2156 | continue; |
2157 | ||
c9de560d AT |
2158 | err = ext4_mb_load_buddy(sb, group, &e4b); |
2159 | if (err) | |
2160 | goto out; | |
2161 | ||
2162 | ext4_lock_group(sb, group); | |
8a57d9d6 CW |
2163 | |
2164 | /* | |
2165 | * We need to check again after locking the | |
2166 | * block group | |
2167 | */ | |
c9de560d | 2168 | if (!ext4_mb_good_group(ac, group, cr)) { |
c9de560d | 2169 | ext4_unlock_group(sb, group); |
e39e07fd | 2170 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
2171 | continue; |
2172 | } | |
2173 | ||
2174 | ac->ac_groups_scanned++; | |
40ae3487 | 2175 | if (cr == 0 && ac->ac_2order < sb->s_blocksize_bits+2) |
c9de560d | 2176 | ext4_mb_simple_scan_group(ac, &e4b); |
506bf2d8 ES |
2177 | else if (cr == 1 && sbi->s_stripe && |
2178 | !(ac->ac_g_ex.fe_len % sbi->s_stripe)) | |
c9de560d AT |
2179 | ext4_mb_scan_aligned(ac, &e4b); |
2180 | else | |
2181 | ext4_mb_complex_scan_group(ac, &e4b); | |
2182 | ||
2183 | ext4_unlock_group(sb, group); | |
e39e07fd | 2184 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
2185 | |
2186 | if (ac->ac_status != AC_STATUS_CONTINUE) | |
2187 | break; | |
2188 | } | |
2189 | } | |
2190 | ||
2191 | if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND && | |
2192 | !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { | |
2193 | /* | |
2194 | * We've been searching too long. Let's try to allocate | |
2195 | * the best chunk we've found so far | |
2196 | */ | |
2197 | ||
2198 | ext4_mb_try_best_found(ac, &e4b); | |
2199 | if (ac->ac_status != AC_STATUS_FOUND) { | |
2200 | /* | |
2201 | * Someone more lucky has already allocated it. | |
2202 | * The only thing we can do is just take first | |
2203 | * found block(s) | |
2204 | printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n"); | |
2205 | */ | |
2206 | ac->ac_b_ex.fe_group = 0; | |
2207 | ac->ac_b_ex.fe_start = 0; | |
2208 | ac->ac_b_ex.fe_len = 0; | |
2209 | ac->ac_status = AC_STATUS_CONTINUE; | |
2210 | ac->ac_flags |= EXT4_MB_HINT_FIRST; | |
2211 | cr = 3; | |
2212 | atomic_inc(&sbi->s_mb_lost_chunks); | |
2213 | goto repeat; | |
2214 | } | |
2215 | } | |
2216 | out: | |
2217 | return err; | |
2218 | } | |
2219 | ||
c9de560d AT |
2220 | static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos) |
2221 | { | |
2222 | struct super_block *sb = seq->private; | |
c9de560d AT |
2223 | ext4_group_t group; |
2224 | ||
8df9675f | 2225 | if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) |
c9de560d | 2226 | return NULL; |
c9de560d | 2227 | group = *pos + 1; |
a9df9a49 | 2228 | return (void *) ((unsigned long) group); |
c9de560d AT |
2229 | } |
2230 | ||
2231 | static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos) | |
2232 | { | |
2233 | struct super_block *sb = seq->private; | |
c9de560d AT |
2234 | ext4_group_t group; |
2235 | ||
2236 | ++*pos; | |
8df9675f | 2237 | if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) |
c9de560d AT |
2238 | return NULL; |
2239 | group = *pos + 1; | |
a9df9a49 | 2240 | return (void *) ((unsigned long) group); |
c9de560d AT |
2241 | } |
2242 | ||
2243 | static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v) | |
2244 | { | |
2245 | struct super_block *sb = seq->private; | |
a9df9a49 | 2246 | ext4_group_t group = (ext4_group_t) ((unsigned long) v); |
c9de560d | 2247 | int i; |
1c8457ca | 2248 | int err, buddy_loaded = 0; |
c9de560d | 2249 | struct ext4_buddy e4b; |
1c8457ca | 2250 | struct ext4_group_info *grinfo; |
c9de560d AT |
2251 | struct sg { |
2252 | struct ext4_group_info info; | |
a36b4498 | 2253 | ext4_grpblk_t counters[16]; |
c9de560d AT |
2254 | } sg; |
2255 | ||
2256 | group--; | |
2257 | if (group == 0) | |
2258 | seq_printf(seq, "#%-5s: %-5s %-5s %-5s " | |
2259 | "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s " | |
2260 | "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n", | |
2261 | "group", "free", "frags", "first", | |
2262 | "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6", | |
2263 | "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13"); | |
2264 | ||
2265 | i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) + | |
2266 | sizeof(struct ext4_group_info); | |
1c8457ca AK |
2267 | grinfo = ext4_get_group_info(sb, group); |
2268 | /* Load the group info in memory only if not already loaded. */ | |
2269 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) { | |
2270 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
2271 | if (err) { | |
2272 | seq_printf(seq, "#%-5u: I/O error\n", group); | |
2273 | return 0; | |
2274 | } | |
2275 | buddy_loaded = 1; | |
c9de560d | 2276 | } |
1c8457ca | 2277 | |
c9de560d | 2278 | memcpy(&sg, ext4_get_group_info(sb, group), i); |
1c8457ca AK |
2279 | |
2280 | if (buddy_loaded) | |
2281 | ext4_mb_unload_buddy(&e4b); | |
c9de560d | 2282 | |
a9df9a49 | 2283 | seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free, |
c9de560d AT |
2284 | sg.info.bb_fragments, sg.info.bb_first_free); |
2285 | for (i = 0; i <= 13; i++) | |
2286 | seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ? | |
2287 | sg.info.bb_counters[i] : 0); | |
2288 | seq_printf(seq, " ]\n"); | |
2289 | ||
2290 | return 0; | |
2291 | } | |
2292 | ||
2293 | static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v) | |
2294 | { | |
2295 | } | |
2296 | ||
7f1346a9 | 2297 | static const struct seq_operations ext4_mb_seq_groups_ops = { |
c9de560d AT |
2298 | .start = ext4_mb_seq_groups_start, |
2299 | .next = ext4_mb_seq_groups_next, | |
2300 | .stop = ext4_mb_seq_groups_stop, | |
2301 | .show = ext4_mb_seq_groups_show, | |
2302 | }; | |
2303 | ||
2304 | static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file) | |
2305 | { | |
d9dda78b | 2306 | struct super_block *sb = PDE_DATA(inode); |
c9de560d AT |
2307 | int rc; |
2308 | ||
2309 | rc = seq_open(file, &ext4_mb_seq_groups_ops); | |
2310 | if (rc == 0) { | |
a271fe85 | 2311 | struct seq_file *m = file->private_data; |
c9de560d AT |
2312 | m->private = sb; |
2313 | } | |
2314 | return rc; | |
2315 | ||
2316 | } | |
2317 | ||
7f1346a9 | 2318 | static const struct file_operations ext4_mb_seq_groups_fops = { |
c9de560d AT |
2319 | .owner = THIS_MODULE, |
2320 | .open = ext4_mb_seq_groups_open, | |
2321 | .read = seq_read, | |
2322 | .llseek = seq_lseek, | |
2323 | .release = seq_release, | |
2324 | }; | |
2325 | ||
fb1813f4 CW |
2326 | static struct kmem_cache *get_groupinfo_cache(int blocksize_bits) |
2327 | { | |
2328 | int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; | |
2329 | struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index]; | |
2330 | ||
2331 | BUG_ON(!cachep); | |
2332 | return cachep; | |
2333 | } | |
5f21b0e6 | 2334 | |
28623c2f TT |
2335 | /* |
2336 | * Allocate the top-level s_group_info array for the specified number | |
2337 | * of groups | |
2338 | */ | |
2339 | int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups) | |
2340 | { | |
2341 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
2342 | unsigned size; | |
2343 | struct ext4_group_info ***new_groupinfo; | |
2344 | ||
2345 | size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >> | |
2346 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
2347 | if (size <= sbi->s_group_info_size) | |
2348 | return 0; | |
2349 | ||
2350 | size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size); | |
2351 | new_groupinfo = ext4_kvzalloc(size, GFP_KERNEL); | |
2352 | if (!new_groupinfo) { | |
2353 | ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group"); | |
2354 | return -ENOMEM; | |
2355 | } | |
2356 | if (sbi->s_group_info) { | |
2357 | memcpy(new_groupinfo, sbi->s_group_info, | |
2358 | sbi->s_group_info_size * sizeof(*sbi->s_group_info)); | |
2359 | ext4_kvfree(sbi->s_group_info); | |
2360 | } | |
2361 | sbi->s_group_info = new_groupinfo; | |
2362 | sbi->s_group_info_size = size / sizeof(*sbi->s_group_info); | |
2363 | ext4_debug("allocated s_groupinfo array for %d meta_bg's\n", | |
2364 | sbi->s_group_info_size); | |
2365 | return 0; | |
2366 | } | |
2367 | ||
5f21b0e6 | 2368 | /* Create and initialize ext4_group_info data for the given group. */ |
920313a7 | 2369 | int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group, |
5f21b0e6 FB |
2370 | struct ext4_group_desc *desc) |
2371 | { | |
fb1813f4 | 2372 | int i; |
5f21b0e6 FB |
2373 | int metalen = 0; |
2374 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
2375 | struct ext4_group_info **meta_group_info; | |
fb1813f4 | 2376 | struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
5f21b0e6 FB |
2377 | |
2378 | /* | |
2379 | * First check if this group is the first of a reserved block. | |
2380 | * If it's true, we have to allocate a new table of pointers | |
2381 | * to ext4_group_info structures | |
2382 | */ | |
2383 | if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { | |
2384 | metalen = sizeof(*meta_group_info) << | |
2385 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
2386 | meta_group_info = kmalloc(metalen, GFP_KERNEL); | |
2387 | if (meta_group_info == NULL) { | |
7f6a11e7 | 2388 | ext4_msg(sb, KERN_ERR, "can't allocate mem " |
9d8b9ec4 | 2389 | "for a buddy group"); |
5f21b0e6 FB |
2390 | goto exit_meta_group_info; |
2391 | } | |
2392 | sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = | |
2393 | meta_group_info; | |
2394 | } | |
2395 | ||
5f21b0e6 FB |
2396 | meta_group_info = |
2397 | sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]; | |
2398 | i = group & (EXT4_DESC_PER_BLOCK(sb) - 1); | |
2399 | ||
85556c9a | 2400 | meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_KERNEL); |
5f21b0e6 | 2401 | if (meta_group_info[i] == NULL) { |
7f6a11e7 | 2402 | ext4_msg(sb, KERN_ERR, "can't allocate buddy mem"); |
5f21b0e6 FB |
2403 | goto exit_group_info; |
2404 | } | |
2405 | set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, | |
2406 | &(meta_group_info[i]->bb_state)); | |
2407 | ||
2408 | /* | |
2409 | * initialize bb_free to be able to skip | |
2410 | * empty groups without initialization | |
2411 | */ | |
2412 | if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { | |
2413 | meta_group_info[i]->bb_free = | |
cff1dfd7 | 2414 | ext4_free_clusters_after_init(sb, group, desc); |
5f21b0e6 FB |
2415 | } else { |
2416 | meta_group_info[i]->bb_free = | |
021b65bb | 2417 | ext4_free_group_clusters(sb, desc); |
5f21b0e6 FB |
2418 | } |
2419 | ||
2420 | INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list); | |
920313a7 | 2421 | init_rwsem(&meta_group_info[i]->alloc_sem); |
64e290ec | 2422 | meta_group_info[i]->bb_free_root = RB_ROOT; |
8a57d9d6 | 2423 | meta_group_info[i]->bb_largest_free_order = -1; /* uninit */ |
5f21b0e6 FB |
2424 | |
2425 | #ifdef DOUBLE_CHECK | |
2426 | { | |
2427 | struct buffer_head *bh; | |
2428 | meta_group_info[i]->bb_bitmap = | |
2429 | kmalloc(sb->s_blocksize, GFP_KERNEL); | |
2430 | BUG_ON(meta_group_info[i]->bb_bitmap == NULL); | |
2431 | bh = ext4_read_block_bitmap(sb, group); | |
2432 | BUG_ON(bh == NULL); | |
2433 | memcpy(meta_group_info[i]->bb_bitmap, bh->b_data, | |
2434 | sb->s_blocksize); | |
2435 | put_bh(bh); | |
2436 | } | |
2437 | #endif | |
2438 | ||
2439 | return 0; | |
2440 | ||
2441 | exit_group_info: | |
2442 | /* If a meta_group_info table has been allocated, release it now */ | |
caaf7a29 | 2443 | if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { |
5f21b0e6 | 2444 | kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]); |
caaf7a29 TM |
2445 | sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL; |
2446 | } | |
5f21b0e6 FB |
2447 | exit_meta_group_info: |
2448 | return -ENOMEM; | |
2449 | } /* ext4_mb_add_groupinfo */ | |
2450 | ||
c9de560d AT |
2451 | static int ext4_mb_init_backend(struct super_block *sb) |
2452 | { | |
8df9675f | 2453 | ext4_group_t ngroups = ext4_get_groups_count(sb); |
c9de560d | 2454 | ext4_group_t i; |
c9de560d | 2455 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
28623c2f | 2456 | int err; |
5f21b0e6 | 2457 | struct ext4_group_desc *desc; |
fb1813f4 | 2458 | struct kmem_cache *cachep; |
5f21b0e6 | 2459 | |
28623c2f TT |
2460 | err = ext4_mb_alloc_groupinfo(sb, ngroups); |
2461 | if (err) | |
2462 | return err; | |
c9de560d | 2463 | |
c9de560d AT |
2464 | sbi->s_buddy_cache = new_inode(sb); |
2465 | if (sbi->s_buddy_cache == NULL) { | |
9d8b9ec4 | 2466 | ext4_msg(sb, KERN_ERR, "can't get new inode"); |
c9de560d AT |
2467 | goto err_freesgi; |
2468 | } | |
48e6061b YJ |
2469 | /* To avoid potentially colliding with an valid on-disk inode number, |
2470 | * use EXT4_BAD_INO for the buddy cache inode number. This inode is | |
2471 | * not in the inode hash, so it should never be found by iget(), but | |
2472 | * this will avoid confusion if it ever shows up during debugging. */ | |
2473 | sbi->s_buddy_cache->i_ino = EXT4_BAD_INO; | |
c9de560d | 2474 | EXT4_I(sbi->s_buddy_cache)->i_disksize = 0; |
8df9675f | 2475 | for (i = 0; i < ngroups; i++) { |
c9de560d AT |
2476 | desc = ext4_get_group_desc(sb, i, NULL); |
2477 | if (desc == NULL) { | |
9d8b9ec4 | 2478 | ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i); |
c9de560d AT |
2479 | goto err_freebuddy; |
2480 | } | |
5f21b0e6 FB |
2481 | if (ext4_mb_add_groupinfo(sb, i, desc) != 0) |
2482 | goto err_freebuddy; | |
c9de560d AT |
2483 | } |
2484 | ||
2485 | return 0; | |
2486 | ||
2487 | err_freebuddy: | |
fb1813f4 | 2488 | cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
f1fa3342 | 2489 | while (i-- > 0) |
fb1813f4 | 2490 | kmem_cache_free(cachep, ext4_get_group_info(sb, i)); |
28623c2f | 2491 | i = sbi->s_group_info_size; |
f1fa3342 | 2492 | while (i-- > 0) |
c9de560d AT |
2493 | kfree(sbi->s_group_info[i]); |
2494 | iput(sbi->s_buddy_cache); | |
2495 | err_freesgi: | |
f18a5f21 | 2496 | ext4_kvfree(sbi->s_group_info); |
c9de560d AT |
2497 | return -ENOMEM; |
2498 | } | |
2499 | ||
2892c15d ES |
2500 | static void ext4_groupinfo_destroy_slabs(void) |
2501 | { | |
2502 | int i; | |
2503 | ||
2504 | for (i = 0; i < NR_GRPINFO_CACHES; i++) { | |
2505 | if (ext4_groupinfo_caches[i]) | |
2506 | kmem_cache_destroy(ext4_groupinfo_caches[i]); | |
2507 | ext4_groupinfo_caches[i] = NULL; | |
2508 | } | |
2509 | } | |
2510 | ||
2511 | static int ext4_groupinfo_create_slab(size_t size) | |
2512 | { | |
2513 | static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex); | |
2514 | int slab_size; | |
2515 | int blocksize_bits = order_base_2(size); | |
2516 | int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; | |
2517 | struct kmem_cache *cachep; | |
2518 | ||
2519 | if (cache_index >= NR_GRPINFO_CACHES) | |
2520 | return -EINVAL; | |
2521 | ||
2522 | if (unlikely(cache_index < 0)) | |
2523 | cache_index = 0; | |
2524 | ||
2525 | mutex_lock(&ext4_grpinfo_slab_create_mutex); | |
2526 | if (ext4_groupinfo_caches[cache_index]) { | |
2527 | mutex_unlock(&ext4_grpinfo_slab_create_mutex); | |
2528 | return 0; /* Already created */ | |
2529 | } | |
2530 | ||
2531 | slab_size = offsetof(struct ext4_group_info, | |
2532 | bb_counters[blocksize_bits + 2]); | |
2533 | ||
2534 | cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index], | |
2535 | slab_size, 0, SLAB_RECLAIM_ACCOUNT, | |
2536 | NULL); | |
2537 | ||
823ba01f TM |
2538 | ext4_groupinfo_caches[cache_index] = cachep; |
2539 | ||
2892c15d ES |
2540 | mutex_unlock(&ext4_grpinfo_slab_create_mutex); |
2541 | if (!cachep) { | |
9d8b9ec4 TT |
2542 | printk(KERN_EMERG |
2543 | "EXT4-fs: no memory for groupinfo slab cache\n"); | |
2892c15d ES |
2544 | return -ENOMEM; |
2545 | } | |
2546 | ||
2892c15d ES |
2547 | return 0; |
2548 | } | |
2549 | ||
9d99012f | 2550 | int ext4_mb_init(struct super_block *sb) |
c9de560d AT |
2551 | { |
2552 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
6be2ded1 | 2553 | unsigned i, j; |
c9de560d AT |
2554 | unsigned offset; |
2555 | unsigned max; | |
74767c5a | 2556 | int ret; |
c9de560d | 2557 | |
1927805e | 2558 | i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets); |
c9de560d AT |
2559 | |
2560 | sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL); | |
2561 | if (sbi->s_mb_offsets == NULL) { | |
fb1813f4 CW |
2562 | ret = -ENOMEM; |
2563 | goto out; | |
c9de560d | 2564 | } |
ff7ef329 | 2565 | |
1927805e | 2566 | i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs); |
c9de560d AT |
2567 | sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL); |
2568 | if (sbi->s_mb_maxs == NULL) { | |
fb1813f4 CW |
2569 | ret = -ENOMEM; |
2570 | goto out; | |
2571 | } | |
2572 | ||
2892c15d ES |
2573 | ret = ext4_groupinfo_create_slab(sb->s_blocksize); |
2574 | if (ret < 0) | |
2575 | goto out; | |
c9de560d AT |
2576 | |
2577 | /* order 0 is regular bitmap */ | |
2578 | sbi->s_mb_maxs[0] = sb->s_blocksize << 3; | |
2579 | sbi->s_mb_offsets[0] = 0; | |
2580 | ||
2581 | i = 1; | |
2582 | offset = 0; | |
2583 | max = sb->s_blocksize << 2; | |
2584 | do { | |
2585 | sbi->s_mb_offsets[i] = offset; | |
2586 | sbi->s_mb_maxs[i] = max; | |
2587 | offset += 1 << (sb->s_blocksize_bits - i); | |
2588 | max = max >> 1; | |
2589 | i++; | |
2590 | } while (i <= sb->s_blocksize_bits + 1); | |
2591 | ||
c9de560d | 2592 | spin_lock_init(&sbi->s_md_lock); |
c9de560d AT |
2593 | spin_lock_init(&sbi->s_bal_lock); |
2594 | ||
2595 | sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN; | |
2596 | sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN; | |
2597 | sbi->s_mb_stats = MB_DEFAULT_STATS; | |
2598 | sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD; | |
2599 | sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS; | |
27baebb8 TT |
2600 | /* |
2601 | * The default group preallocation is 512, which for 4k block | |
2602 | * sizes translates to 2 megabytes. However for bigalloc file | |
2603 | * systems, this is probably too big (i.e, if the cluster size | |
2604 | * is 1 megabyte, then group preallocation size becomes half a | |
2605 | * gigabyte!). As a default, we will keep a two megabyte | |
2606 | * group pralloc size for cluster sizes up to 64k, and after | |
2607 | * that, we will force a minimum group preallocation size of | |
2608 | * 32 clusters. This translates to 8 megs when the cluster | |
2609 | * size is 256k, and 32 megs when the cluster size is 1 meg, | |
2610 | * which seems reasonable as a default. | |
2611 | */ | |
2612 | sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >> | |
2613 | sbi->s_cluster_bits, 32); | |
d7a1fee1 DE |
2614 | /* |
2615 | * If there is a s_stripe > 1, then we set the s_mb_group_prealloc | |
2616 | * to the lowest multiple of s_stripe which is bigger than | |
2617 | * the s_mb_group_prealloc as determined above. We want | |
2618 | * the preallocation size to be an exact multiple of the | |
2619 | * RAID stripe size so that preallocations don't fragment | |
2620 | * the stripes. | |
2621 | */ | |
2622 | if (sbi->s_stripe > 1) { | |
2623 | sbi->s_mb_group_prealloc = roundup( | |
2624 | sbi->s_mb_group_prealloc, sbi->s_stripe); | |
2625 | } | |
c9de560d | 2626 | |
730c213c | 2627 | sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group); |
c9de560d | 2628 | if (sbi->s_locality_groups == NULL) { |
fb1813f4 | 2629 | ret = -ENOMEM; |
029b10c5 | 2630 | goto out; |
c9de560d | 2631 | } |
730c213c | 2632 | for_each_possible_cpu(i) { |
c9de560d | 2633 | struct ext4_locality_group *lg; |
730c213c | 2634 | lg = per_cpu_ptr(sbi->s_locality_groups, i); |
c9de560d | 2635 | mutex_init(&lg->lg_mutex); |
6be2ded1 AK |
2636 | for (j = 0; j < PREALLOC_TB_SIZE; j++) |
2637 | INIT_LIST_HEAD(&lg->lg_prealloc_list[j]); | |
c9de560d AT |
2638 | spin_lock_init(&lg->lg_prealloc_lock); |
2639 | } | |
2640 | ||
79a77c5a YJ |
2641 | /* init file for buddy data */ |
2642 | ret = ext4_mb_init_backend(sb); | |
7aa0baea TM |
2643 | if (ret != 0) |
2644 | goto out_free_locality_groups; | |
79a77c5a | 2645 | |
296c355c TT |
2646 | if (sbi->s_proc) |
2647 | proc_create_data("mb_groups", S_IRUGO, sbi->s_proc, | |
2648 | &ext4_mb_seq_groups_fops, sb); | |
c9de560d | 2649 | |
7aa0baea TM |
2650 | return 0; |
2651 | ||
2652 | out_free_locality_groups: | |
2653 | free_percpu(sbi->s_locality_groups); | |
2654 | sbi->s_locality_groups = NULL; | |
fb1813f4 | 2655 | out: |
7aa0baea TM |
2656 | kfree(sbi->s_mb_offsets); |
2657 | sbi->s_mb_offsets = NULL; | |
2658 | kfree(sbi->s_mb_maxs); | |
2659 | sbi->s_mb_maxs = NULL; | |
fb1813f4 | 2660 | return ret; |
c9de560d AT |
2661 | } |
2662 | ||
955ce5f5 | 2663 | /* need to called with the ext4 group lock held */ |
c9de560d AT |
2664 | static void ext4_mb_cleanup_pa(struct ext4_group_info *grp) |
2665 | { | |
2666 | struct ext4_prealloc_space *pa; | |
2667 | struct list_head *cur, *tmp; | |
2668 | int count = 0; | |
2669 | ||
2670 | list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) { | |
2671 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); | |
2672 | list_del(&pa->pa_group_list); | |
2673 | count++; | |
688f05a0 | 2674 | kmem_cache_free(ext4_pspace_cachep, pa); |
c9de560d AT |
2675 | } |
2676 | if (count) | |
6ba495e9 | 2677 | mb_debug(1, "mballoc: %u PAs left\n", count); |
c9de560d AT |
2678 | |
2679 | } | |
2680 | ||
2681 | int ext4_mb_release(struct super_block *sb) | |
2682 | { | |
8df9675f | 2683 | ext4_group_t ngroups = ext4_get_groups_count(sb); |
c9de560d AT |
2684 | ext4_group_t i; |
2685 | int num_meta_group_infos; | |
2686 | struct ext4_group_info *grinfo; | |
2687 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
fb1813f4 | 2688 | struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); |
c9de560d | 2689 | |
95599968 SQ |
2690 | if (sbi->s_proc) |
2691 | remove_proc_entry("mb_groups", sbi->s_proc); | |
2692 | ||
c9de560d | 2693 | if (sbi->s_group_info) { |
8df9675f | 2694 | for (i = 0; i < ngroups; i++) { |
c9de560d AT |
2695 | grinfo = ext4_get_group_info(sb, i); |
2696 | #ifdef DOUBLE_CHECK | |
2697 | kfree(grinfo->bb_bitmap); | |
2698 | #endif | |
2699 | ext4_lock_group(sb, i); | |
2700 | ext4_mb_cleanup_pa(grinfo); | |
2701 | ext4_unlock_group(sb, i); | |
fb1813f4 | 2702 | kmem_cache_free(cachep, grinfo); |
c9de560d | 2703 | } |
8df9675f | 2704 | num_meta_group_infos = (ngroups + |
c9de560d AT |
2705 | EXT4_DESC_PER_BLOCK(sb) - 1) >> |
2706 | EXT4_DESC_PER_BLOCK_BITS(sb); | |
2707 | for (i = 0; i < num_meta_group_infos; i++) | |
2708 | kfree(sbi->s_group_info[i]); | |
f18a5f21 | 2709 | ext4_kvfree(sbi->s_group_info); |
c9de560d AT |
2710 | } |
2711 | kfree(sbi->s_mb_offsets); | |
2712 | kfree(sbi->s_mb_maxs); | |
2713 | if (sbi->s_buddy_cache) | |
2714 | iput(sbi->s_buddy_cache); | |
2715 | if (sbi->s_mb_stats) { | |
9d8b9ec4 TT |
2716 | ext4_msg(sb, KERN_INFO, |
2717 | "mballoc: %u blocks %u reqs (%u success)", | |
c9de560d AT |
2718 | atomic_read(&sbi->s_bal_allocated), |
2719 | atomic_read(&sbi->s_bal_reqs), | |
2720 | atomic_read(&sbi->s_bal_success)); | |
9d8b9ec4 TT |
2721 | ext4_msg(sb, KERN_INFO, |
2722 | "mballoc: %u extents scanned, %u goal hits, " | |
2723 | "%u 2^N hits, %u breaks, %u lost", | |
c9de560d AT |
2724 | atomic_read(&sbi->s_bal_ex_scanned), |
2725 | atomic_read(&sbi->s_bal_goals), | |
2726 | atomic_read(&sbi->s_bal_2orders), | |
2727 | atomic_read(&sbi->s_bal_breaks), | |
2728 | atomic_read(&sbi->s_mb_lost_chunks)); | |
9d8b9ec4 TT |
2729 | ext4_msg(sb, KERN_INFO, |
2730 | "mballoc: %lu generated and it took %Lu", | |
ced156e4 | 2731 | sbi->s_mb_buddies_generated, |
c9de560d | 2732 | sbi->s_mb_generation_time); |
9d8b9ec4 TT |
2733 | ext4_msg(sb, KERN_INFO, |
2734 | "mballoc: %u preallocated, %u discarded", | |
c9de560d AT |
2735 | atomic_read(&sbi->s_mb_preallocated), |
2736 | atomic_read(&sbi->s_mb_discarded)); | |
2737 | } | |
2738 | ||
730c213c | 2739 | free_percpu(sbi->s_locality_groups); |
c9de560d AT |
2740 | |
2741 | return 0; | |
2742 | } | |
2743 | ||
77ca6cdf | 2744 | static inline int ext4_issue_discard(struct super_block *sb, |
84130193 | 2745 | ext4_group_t block_group, ext4_grpblk_t cluster, int count) |
5c521830 | 2746 | { |
5c521830 JZ |
2747 | ext4_fsblk_t discard_block; |
2748 | ||
84130193 TT |
2749 | discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) + |
2750 | ext4_group_first_block_no(sb, block_group)); | |
2751 | count = EXT4_C2B(EXT4_SB(sb), count); | |
5c521830 JZ |
2752 | trace_ext4_discard_blocks(sb, |
2753 | (unsigned long long) discard_block, count); | |
93259636 | 2754 | return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0); |
5c521830 JZ |
2755 | } |
2756 | ||
3e624fc7 TT |
2757 | /* |
2758 | * This function is called by the jbd2 layer once the commit has finished, | |
2759 | * so we know we can free the blocks that were released with that commit. | |
2760 | */ | |
18aadd47 BJ |
2761 | static void ext4_free_data_callback(struct super_block *sb, |
2762 | struct ext4_journal_cb_entry *jce, | |
2763 | int rc) | |
c9de560d | 2764 | { |
18aadd47 | 2765 | struct ext4_free_data *entry = (struct ext4_free_data *)jce; |
c9de560d | 2766 | struct ext4_buddy e4b; |
c894058d | 2767 | struct ext4_group_info *db; |
d9f34504 | 2768 | int err, count = 0, count2 = 0; |
c9de560d | 2769 | |
18aadd47 BJ |
2770 | mb_debug(1, "gonna free %u blocks in group %u (0x%p):", |
2771 | entry->efd_count, entry->efd_group, entry); | |
c9de560d | 2772 | |
d71c1ae2 LC |
2773 | if (test_opt(sb, DISCARD)) { |
2774 | err = ext4_issue_discard(sb, entry->efd_group, | |
2775 | entry->efd_start_cluster, | |
2776 | entry->efd_count); | |
2777 | if (err && err != -EOPNOTSUPP) | |
2778 | ext4_msg(sb, KERN_WARNING, "discard request in" | |
2779 | " group:%d block:%d count:%d failed" | |
2780 | " with %d", entry->efd_group, | |
2781 | entry->efd_start_cluster, | |
2782 | entry->efd_count, err); | |
2783 | } | |
c9de560d | 2784 | |
18aadd47 BJ |
2785 | err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b); |
2786 | /* we expect to find existing buddy because it's pinned */ | |
2787 | BUG_ON(err != 0); | |
b90f6870 | 2788 | |
c9de560d | 2789 | |
18aadd47 BJ |
2790 | db = e4b.bd_info; |
2791 | /* there are blocks to put in buddy to make them really free */ | |
2792 | count += entry->efd_count; | |
2793 | count2++; | |
2794 | ext4_lock_group(sb, entry->efd_group); | |
2795 | /* Take it out of per group rb tree */ | |
2796 | rb_erase(&entry->efd_node, &(db->bb_free_root)); | |
2797 | mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count); | |
c894058d | 2798 | |
18aadd47 BJ |
2799 | /* |
2800 | * Clear the trimmed flag for the group so that the next | |
2801 | * ext4_trim_fs can trim it. | |
2802 | * If the volume is mounted with -o discard, online discard | |
2803 | * is supported and the free blocks will be trimmed online. | |
2804 | */ | |
2805 | if (!test_opt(sb, DISCARD)) | |
2806 | EXT4_MB_GRP_CLEAR_TRIMMED(db); | |
3d56b8d2 | 2807 | |
18aadd47 BJ |
2808 | if (!db->bb_free_root.rb_node) { |
2809 | /* No more items in the per group rb tree | |
2810 | * balance refcounts from ext4_mb_free_metadata() | |
2811 | */ | |
2812 | page_cache_release(e4b.bd_buddy_page); | |
2813 | page_cache_release(e4b.bd_bitmap_page); | |
3e624fc7 | 2814 | } |
18aadd47 BJ |
2815 | ext4_unlock_group(sb, entry->efd_group); |
2816 | kmem_cache_free(ext4_free_data_cachep, entry); | |
2817 | ext4_mb_unload_buddy(&e4b); | |
c9de560d | 2818 | |
6ba495e9 | 2819 | mb_debug(1, "freed %u blocks in %u structures\n", count, count2); |
c9de560d AT |
2820 | } |
2821 | ||
5dabfc78 | 2822 | int __init ext4_init_mballoc(void) |
c9de560d | 2823 | { |
16828088 TT |
2824 | ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space, |
2825 | SLAB_RECLAIM_ACCOUNT); | |
c9de560d AT |
2826 | if (ext4_pspace_cachep == NULL) |
2827 | return -ENOMEM; | |
2828 | ||
16828088 TT |
2829 | ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context, |
2830 | SLAB_RECLAIM_ACCOUNT); | |
256bdb49 ES |
2831 | if (ext4_ac_cachep == NULL) { |
2832 | kmem_cache_destroy(ext4_pspace_cachep); | |
2833 | return -ENOMEM; | |
2834 | } | |
c894058d | 2835 | |
18aadd47 BJ |
2836 | ext4_free_data_cachep = KMEM_CACHE(ext4_free_data, |
2837 | SLAB_RECLAIM_ACCOUNT); | |
2838 | if (ext4_free_data_cachep == NULL) { | |
c894058d AK |
2839 | kmem_cache_destroy(ext4_pspace_cachep); |
2840 | kmem_cache_destroy(ext4_ac_cachep); | |
2841 | return -ENOMEM; | |
2842 | } | |
c9de560d AT |
2843 | return 0; |
2844 | } | |
2845 | ||
5dabfc78 | 2846 | void ext4_exit_mballoc(void) |
c9de560d | 2847 | { |
60e6679e | 2848 | /* |
3e03f9ca JDB |
2849 | * Wait for completion of call_rcu()'s on ext4_pspace_cachep |
2850 | * before destroying the slab cache. | |
2851 | */ | |
2852 | rcu_barrier(); | |
c9de560d | 2853 | kmem_cache_destroy(ext4_pspace_cachep); |
256bdb49 | 2854 | kmem_cache_destroy(ext4_ac_cachep); |
18aadd47 | 2855 | kmem_cache_destroy(ext4_free_data_cachep); |
2892c15d | 2856 | ext4_groupinfo_destroy_slabs(); |
c9de560d AT |
2857 | } |
2858 | ||
2859 | ||
2860 | /* | |
73b2c716 | 2861 | * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps |
c9de560d AT |
2862 | * Returns 0 if success or error code |
2863 | */ | |
4ddfef7b ES |
2864 | static noinline_for_stack int |
2865 | ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac, | |
53accfa9 | 2866 | handle_t *handle, unsigned int reserv_clstrs) |
c9de560d AT |
2867 | { |
2868 | struct buffer_head *bitmap_bh = NULL; | |
c9de560d AT |
2869 | struct ext4_group_desc *gdp; |
2870 | struct buffer_head *gdp_bh; | |
2871 | struct ext4_sb_info *sbi; | |
2872 | struct super_block *sb; | |
2873 | ext4_fsblk_t block; | |
519deca0 | 2874 | int err, len; |
c9de560d AT |
2875 | |
2876 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
2877 | BUG_ON(ac->ac_b_ex.fe_len <= 0); | |
2878 | ||
2879 | sb = ac->ac_sb; | |
2880 | sbi = EXT4_SB(sb); | |
c9de560d AT |
2881 | |
2882 | err = -EIO; | |
574ca174 | 2883 | bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group); |
c9de560d AT |
2884 | if (!bitmap_bh) |
2885 | goto out_err; | |
2886 | ||
5d601255 | 2887 | BUFFER_TRACE(bitmap_bh, "getting write access"); |
c9de560d AT |
2888 | err = ext4_journal_get_write_access(handle, bitmap_bh); |
2889 | if (err) | |
2890 | goto out_err; | |
2891 | ||
2892 | err = -EIO; | |
2893 | gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh); | |
2894 | if (!gdp) | |
2895 | goto out_err; | |
2896 | ||
a9df9a49 | 2897 | ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group, |
021b65bb | 2898 | ext4_free_group_clusters(sb, gdp)); |
03cddb80 | 2899 | |
5d601255 | 2900 | BUFFER_TRACE(gdp_bh, "get_write_access"); |
c9de560d AT |
2901 | err = ext4_journal_get_write_access(handle, gdp_bh); |
2902 | if (err) | |
2903 | goto out_err; | |
2904 | ||
bda00de7 | 2905 | block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); |
c9de560d | 2906 | |
53accfa9 | 2907 | len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len); |
6fd058f7 | 2908 | if (!ext4_data_block_valid(sbi, block, len)) { |
12062ddd | 2909 | ext4_error(sb, "Allocating blocks %llu-%llu which overlap " |
1084f252 | 2910 | "fs metadata", block, block+len); |
519deca0 AK |
2911 | /* File system mounted not to panic on error |
2912 | * Fix the bitmap and repeat the block allocation | |
2913 | * We leak some of the blocks here. | |
2914 | */ | |
955ce5f5 | 2915 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); |
c3e94d1d YY |
2916 | ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, |
2917 | ac->ac_b_ex.fe_len); | |
955ce5f5 | 2918 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); |
0390131b | 2919 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); |
519deca0 AK |
2920 | if (!err) |
2921 | err = -EAGAIN; | |
2922 | goto out_err; | |
c9de560d | 2923 | } |
955ce5f5 AK |
2924 | |
2925 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
c9de560d AT |
2926 | #ifdef AGGRESSIVE_CHECK |
2927 | { | |
2928 | int i; | |
2929 | for (i = 0; i < ac->ac_b_ex.fe_len; i++) { | |
2930 | BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i, | |
2931 | bitmap_bh->b_data)); | |
2932 | } | |
2933 | } | |
2934 | #endif | |
c3e94d1d YY |
2935 | ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, |
2936 | ac->ac_b_ex.fe_len); | |
c9de560d AT |
2937 | if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { |
2938 | gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); | |
021b65bb | 2939 | ext4_free_group_clusters_set(sb, gdp, |
cff1dfd7 | 2940 | ext4_free_clusters_after_init(sb, |
021b65bb | 2941 | ac->ac_b_ex.fe_group, gdp)); |
c9de560d | 2942 | } |
021b65bb TT |
2943 | len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len; |
2944 | ext4_free_group_clusters_set(sb, gdp, len); | |
79f1ba49 | 2945 | ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh); |
feb0ab32 | 2946 | ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp); |
955ce5f5 AK |
2947 | |
2948 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
57042651 | 2949 | percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len); |
d2a17637 | 2950 | /* |
6bc6e63f | 2951 | * Now reduce the dirty block count also. Should not go negative |
d2a17637 | 2952 | */ |
6bc6e63f AK |
2953 | if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED)) |
2954 | /* release all the reserved blocks if non delalloc */ | |
57042651 TT |
2955 | percpu_counter_sub(&sbi->s_dirtyclusters_counter, |
2956 | reserv_clstrs); | |
c9de560d | 2957 | |
772cb7c8 JS |
2958 | if (sbi->s_log_groups_per_flex) { |
2959 | ext4_group_t flex_group = ext4_flex_group(sbi, | |
2960 | ac->ac_b_ex.fe_group); | |
90ba983f TT |
2961 | atomic64_sub(ac->ac_b_ex.fe_len, |
2962 | &sbi->s_flex_groups[flex_group].free_clusters); | |
772cb7c8 JS |
2963 | } |
2964 | ||
0390131b | 2965 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); |
c9de560d AT |
2966 | if (err) |
2967 | goto out_err; | |
0390131b | 2968 | err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh); |
c9de560d AT |
2969 | |
2970 | out_err: | |
42a10add | 2971 | brelse(bitmap_bh); |
c9de560d AT |
2972 | return err; |
2973 | } | |
2974 | ||
2975 | /* | |
2976 | * here we normalize request for locality group | |
d7a1fee1 DE |
2977 | * Group request are normalized to s_mb_group_prealloc, which goes to |
2978 | * s_strip if we set the same via mount option. | |
2979 | * s_mb_group_prealloc can be configured via | |
b713a5ec | 2980 | * /sys/fs/ext4/<partition>/mb_group_prealloc |
c9de560d AT |
2981 | * |
2982 | * XXX: should we try to preallocate more than the group has now? | |
2983 | */ | |
2984 | static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac) | |
2985 | { | |
2986 | struct super_block *sb = ac->ac_sb; | |
2987 | struct ext4_locality_group *lg = ac->ac_lg; | |
2988 | ||
2989 | BUG_ON(lg == NULL); | |
d7a1fee1 | 2990 | ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc; |
6ba495e9 | 2991 | mb_debug(1, "#%u: goal %u blocks for locality group\n", |
c9de560d AT |
2992 | current->pid, ac->ac_g_ex.fe_len); |
2993 | } | |
2994 | ||
2995 | /* | |
2996 | * Normalization means making request better in terms of | |
2997 | * size and alignment | |
2998 | */ | |
4ddfef7b ES |
2999 | static noinline_for_stack void |
3000 | ext4_mb_normalize_request(struct ext4_allocation_context *ac, | |
c9de560d AT |
3001 | struct ext4_allocation_request *ar) |
3002 | { | |
53accfa9 | 3003 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
c9de560d AT |
3004 | int bsbits, max; |
3005 | ext4_lblk_t end; | |
1592d2c5 CW |
3006 | loff_t size, start_off; |
3007 | loff_t orig_size __maybe_unused; | |
5a0790c2 | 3008 | ext4_lblk_t start; |
c9de560d | 3009 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); |
9a0762c5 | 3010 | struct ext4_prealloc_space *pa; |
c9de560d AT |
3011 | |
3012 | /* do normalize only data requests, metadata requests | |
3013 | do not need preallocation */ | |
3014 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
3015 | return; | |
3016 | ||
3017 | /* sometime caller may want exact blocks */ | |
3018 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) | |
3019 | return; | |
3020 | ||
3021 | /* caller may indicate that preallocation isn't | |
3022 | * required (it's a tail, for example) */ | |
3023 | if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC) | |
3024 | return; | |
3025 | ||
3026 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) { | |
3027 | ext4_mb_normalize_group_request(ac); | |
3028 | return ; | |
3029 | } | |
3030 | ||
3031 | bsbits = ac->ac_sb->s_blocksize_bits; | |
3032 | ||
3033 | /* first, let's learn actual file size | |
3034 | * given current request is allocated */ | |
53accfa9 | 3035 | size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len); |
c9de560d AT |
3036 | size = size << bsbits; |
3037 | if (size < i_size_read(ac->ac_inode)) | |
3038 | size = i_size_read(ac->ac_inode); | |
5a0790c2 | 3039 | orig_size = size; |
c9de560d | 3040 | |
1930479c VC |
3041 | /* max size of free chunks */ |
3042 | max = 2 << bsbits; | |
c9de560d | 3043 | |
1930479c VC |
3044 | #define NRL_CHECK_SIZE(req, size, max, chunk_size) \ |
3045 | (req <= (size) || max <= (chunk_size)) | |
c9de560d AT |
3046 | |
3047 | /* first, try to predict filesize */ | |
3048 | /* XXX: should this table be tunable? */ | |
3049 | start_off = 0; | |
3050 | if (size <= 16 * 1024) { | |
3051 | size = 16 * 1024; | |
3052 | } else if (size <= 32 * 1024) { | |
3053 | size = 32 * 1024; | |
3054 | } else if (size <= 64 * 1024) { | |
3055 | size = 64 * 1024; | |
3056 | } else if (size <= 128 * 1024) { | |
3057 | size = 128 * 1024; | |
3058 | } else if (size <= 256 * 1024) { | |
3059 | size = 256 * 1024; | |
3060 | } else if (size <= 512 * 1024) { | |
3061 | size = 512 * 1024; | |
3062 | } else if (size <= 1024 * 1024) { | |
3063 | size = 1024 * 1024; | |
1930479c | 3064 | } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) { |
c9de560d | 3065 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
1930479c VC |
3066 | (21 - bsbits)) << 21; |
3067 | size = 2 * 1024 * 1024; | |
3068 | } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) { | |
c9de560d AT |
3069 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
3070 | (22 - bsbits)) << 22; | |
3071 | size = 4 * 1024 * 1024; | |
3072 | } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len, | |
1930479c | 3073 | (8<<20)>>bsbits, max, 8 * 1024)) { |
c9de560d AT |
3074 | start_off = ((loff_t)ac->ac_o_ex.fe_logical >> |
3075 | (23 - bsbits)) << 23; | |
3076 | size = 8 * 1024 * 1024; | |
3077 | } else { | |
b27b1535 XW |
3078 | start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits; |
3079 | size = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb), | |
3080 | ac->ac_o_ex.fe_len) << bsbits; | |
c9de560d | 3081 | } |
5a0790c2 AK |
3082 | size = size >> bsbits; |
3083 | start = start_off >> bsbits; | |
c9de560d AT |
3084 | |
3085 | /* don't cover already allocated blocks in selected range */ | |
3086 | if (ar->pleft && start <= ar->lleft) { | |
3087 | size -= ar->lleft + 1 - start; | |
3088 | start = ar->lleft + 1; | |
3089 | } | |
3090 | if (ar->pright && start + size - 1 >= ar->lright) | |
3091 | size -= start + size - ar->lright; | |
3092 | ||
3093 | end = start + size; | |
3094 | ||
3095 | /* check we don't cross already preallocated blocks */ | |
3096 | rcu_read_lock(); | |
9a0762c5 | 3097 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
498e5f24 | 3098 | ext4_lblk_t pa_end; |
c9de560d | 3099 | |
c9de560d AT |
3100 | if (pa->pa_deleted) |
3101 | continue; | |
3102 | spin_lock(&pa->pa_lock); | |
3103 | if (pa->pa_deleted) { | |
3104 | spin_unlock(&pa->pa_lock); | |
3105 | continue; | |
3106 | } | |
3107 | ||
53accfa9 TT |
3108 | pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb), |
3109 | pa->pa_len); | |
c9de560d AT |
3110 | |
3111 | /* PA must not overlap original request */ | |
3112 | BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end || | |
3113 | ac->ac_o_ex.fe_logical < pa->pa_lstart)); | |
3114 | ||
38877f4e ES |
3115 | /* skip PAs this normalized request doesn't overlap with */ |
3116 | if (pa->pa_lstart >= end || pa_end <= start) { | |
c9de560d AT |
3117 | spin_unlock(&pa->pa_lock); |
3118 | continue; | |
3119 | } | |
3120 | BUG_ON(pa->pa_lstart <= start && pa_end >= end); | |
3121 | ||
38877f4e | 3122 | /* adjust start or end to be adjacent to this pa */ |
c9de560d AT |
3123 | if (pa_end <= ac->ac_o_ex.fe_logical) { |
3124 | BUG_ON(pa_end < start); | |
3125 | start = pa_end; | |
38877f4e | 3126 | } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) { |
c9de560d AT |
3127 | BUG_ON(pa->pa_lstart > end); |
3128 | end = pa->pa_lstart; | |
3129 | } | |
3130 | spin_unlock(&pa->pa_lock); | |
3131 | } | |
3132 | rcu_read_unlock(); | |
3133 | size = end - start; | |
3134 | ||
3135 | /* XXX: extra loop to check we really don't overlap preallocations */ | |
3136 | rcu_read_lock(); | |
9a0762c5 | 3137 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
498e5f24 | 3138 | ext4_lblk_t pa_end; |
53accfa9 | 3139 | |
c9de560d AT |
3140 | spin_lock(&pa->pa_lock); |
3141 | if (pa->pa_deleted == 0) { | |
53accfa9 TT |
3142 | pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb), |
3143 | pa->pa_len); | |
c9de560d AT |
3144 | BUG_ON(!(start >= pa_end || end <= pa->pa_lstart)); |
3145 | } | |
3146 | spin_unlock(&pa->pa_lock); | |
3147 | } | |
3148 | rcu_read_unlock(); | |
3149 | ||
3150 | if (start + size <= ac->ac_o_ex.fe_logical && | |
3151 | start > ac->ac_o_ex.fe_logical) { | |
9d8b9ec4 TT |
3152 | ext4_msg(ac->ac_sb, KERN_ERR, |
3153 | "start %lu, size %lu, fe_logical %lu", | |
3154 | (unsigned long) start, (unsigned long) size, | |
3155 | (unsigned long) ac->ac_o_ex.fe_logical); | |
c9de560d AT |
3156 | } |
3157 | BUG_ON(start + size <= ac->ac_o_ex.fe_logical && | |
3158 | start > ac->ac_o_ex.fe_logical); | |
b5b60778 | 3159 | BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); |
c9de560d AT |
3160 | |
3161 | /* now prepare goal request */ | |
3162 | ||
3163 | /* XXX: is it better to align blocks WRT to logical | |
3164 | * placement or satisfy big request as is */ | |
3165 | ac->ac_g_ex.fe_logical = start; | |
53accfa9 | 3166 | ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size); |
c9de560d AT |
3167 | |
3168 | /* define goal start in order to merge */ | |
3169 | if (ar->pright && (ar->lright == (start + size))) { | |
3170 | /* merge to the right */ | |
3171 | ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size, | |
3172 | &ac->ac_f_ex.fe_group, | |
3173 | &ac->ac_f_ex.fe_start); | |
3174 | ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; | |
3175 | } | |
3176 | if (ar->pleft && (ar->lleft + 1 == start)) { | |
3177 | /* merge to the left */ | |
3178 | ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1, | |
3179 | &ac->ac_f_ex.fe_group, | |
3180 | &ac->ac_f_ex.fe_start); | |
3181 | ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; | |
3182 | } | |
3183 | ||
6ba495e9 | 3184 | mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size, |
c9de560d AT |
3185 | (unsigned) orig_size, (unsigned) start); |
3186 | } | |
3187 | ||
3188 | static void ext4_mb_collect_stats(struct ext4_allocation_context *ac) | |
3189 | { | |
3190 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
3191 | ||
3192 | if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) { | |
3193 | atomic_inc(&sbi->s_bal_reqs); | |
3194 | atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated); | |
291dae47 | 3195 | if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len) |
c9de560d AT |
3196 | atomic_inc(&sbi->s_bal_success); |
3197 | atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned); | |
3198 | if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start && | |
3199 | ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group) | |
3200 | atomic_inc(&sbi->s_bal_goals); | |
3201 | if (ac->ac_found > sbi->s_mb_max_to_scan) | |
3202 | atomic_inc(&sbi->s_bal_breaks); | |
3203 | } | |
3204 | ||
296c355c TT |
3205 | if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) |
3206 | trace_ext4_mballoc_alloc(ac); | |
3207 | else | |
3208 | trace_ext4_mballoc_prealloc(ac); | |
c9de560d AT |
3209 | } |
3210 | ||
b844167e CW |
3211 | /* |
3212 | * Called on failure; free up any blocks from the inode PA for this | |
3213 | * context. We don't need this for MB_GROUP_PA because we only change | |
3214 | * pa_free in ext4_mb_release_context(), but on failure, we've already | |
3215 | * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed. | |
3216 | */ | |
3217 | static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac) | |
3218 | { | |
3219 | struct ext4_prealloc_space *pa = ac->ac_pa; | |
86f0afd4 TT |
3220 | struct ext4_buddy e4b; |
3221 | int err; | |
b844167e | 3222 | |
86f0afd4 TT |
3223 | if (pa == NULL) { |
3224 | err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b); | |
3225 | if (err) { | |
3226 | /* | |
3227 | * This should never happen since we pin the | |
3228 | * pages in the ext4_allocation_context so | |
3229 | * ext4_mb_load_buddy() should never fail. | |
3230 | */ | |
3231 | WARN(1, "mb_load_buddy failed (%d)", err); | |
3232 | return; | |
3233 | } | |
3234 | ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group); | |
3235 | mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start, | |
3236 | ac->ac_f_ex.fe_len); | |
3237 | ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group); | |
3238 | return; | |
3239 | } | |
3240 | if (pa->pa_type == MB_INODE_PA) | |
400db9d3 | 3241 | pa->pa_free += ac->ac_b_ex.fe_len; |
b844167e CW |
3242 | } |
3243 | ||
c9de560d AT |
3244 | /* |
3245 | * use blocks preallocated to inode | |
3246 | */ | |
3247 | static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac, | |
3248 | struct ext4_prealloc_space *pa) | |
3249 | { | |
53accfa9 | 3250 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
c9de560d AT |
3251 | ext4_fsblk_t start; |
3252 | ext4_fsblk_t end; | |
3253 | int len; | |
3254 | ||
3255 | /* found preallocated blocks, use them */ | |
3256 | start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart); | |
53accfa9 TT |
3257 | end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len), |
3258 | start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len)); | |
3259 | len = EXT4_NUM_B2C(sbi, end - start); | |
c9de560d AT |
3260 | ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group, |
3261 | &ac->ac_b_ex.fe_start); | |
3262 | ac->ac_b_ex.fe_len = len; | |
3263 | ac->ac_status = AC_STATUS_FOUND; | |
3264 | ac->ac_pa = pa; | |
3265 | ||
3266 | BUG_ON(start < pa->pa_pstart); | |
53accfa9 | 3267 | BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len)); |
c9de560d AT |
3268 | BUG_ON(pa->pa_free < len); |
3269 | pa->pa_free -= len; | |
3270 | ||
6ba495e9 | 3271 | mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa); |
c9de560d AT |
3272 | } |
3273 | ||
3274 | /* | |
3275 | * use blocks preallocated to locality group | |
3276 | */ | |
3277 | static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac, | |
3278 | struct ext4_prealloc_space *pa) | |
3279 | { | |
03cddb80 | 3280 | unsigned int len = ac->ac_o_ex.fe_len; |
6be2ded1 | 3281 | |
c9de560d AT |
3282 | ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart, |
3283 | &ac->ac_b_ex.fe_group, | |
3284 | &ac->ac_b_ex.fe_start); | |
3285 | ac->ac_b_ex.fe_len = len; | |
3286 | ac->ac_status = AC_STATUS_FOUND; | |
3287 | ac->ac_pa = pa; | |
3288 | ||
3289 | /* we don't correct pa_pstart or pa_plen here to avoid | |
26346ff6 | 3290 | * possible race when the group is being loaded concurrently |
c9de560d | 3291 | * instead we correct pa later, after blocks are marked |
26346ff6 AK |
3292 | * in on-disk bitmap -- see ext4_mb_release_context() |
3293 | * Other CPUs are prevented from allocating from this pa by lg_mutex | |
c9de560d | 3294 | */ |
6ba495e9 | 3295 | mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa); |
c9de560d AT |
3296 | } |
3297 | ||
5e745b04 AK |
3298 | /* |
3299 | * Return the prealloc space that have minimal distance | |
3300 | * from the goal block. @cpa is the prealloc | |
3301 | * space that is having currently known minimal distance | |
3302 | * from the goal block. | |
3303 | */ | |
3304 | static struct ext4_prealloc_space * | |
3305 | ext4_mb_check_group_pa(ext4_fsblk_t goal_block, | |
3306 | struct ext4_prealloc_space *pa, | |
3307 | struct ext4_prealloc_space *cpa) | |
3308 | { | |
3309 | ext4_fsblk_t cur_distance, new_distance; | |
3310 | ||
3311 | if (cpa == NULL) { | |
3312 | atomic_inc(&pa->pa_count); | |
3313 | return pa; | |
3314 | } | |
3315 | cur_distance = abs(goal_block - cpa->pa_pstart); | |
3316 | new_distance = abs(goal_block - pa->pa_pstart); | |
3317 | ||
5a54b2f1 | 3318 | if (cur_distance <= new_distance) |
5e745b04 AK |
3319 | return cpa; |
3320 | ||
3321 | /* drop the previous reference */ | |
3322 | atomic_dec(&cpa->pa_count); | |
3323 | atomic_inc(&pa->pa_count); | |
3324 | return pa; | |
3325 | } | |
3326 | ||
c9de560d AT |
3327 | /* |
3328 | * search goal blocks in preallocated space | |
3329 | */ | |
4ddfef7b ES |
3330 | static noinline_for_stack int |
3331 | ext4_mb_use_preallocated(struct ext4_allocation_context *ac) | |
c9de560d | 3332 | { |
53accfa9 | 3333 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
6be2ded1 | 3334 | int order, i; |
c9de560d AT |
3335 | struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); |
3336 | struct ext4_locality_group *lg; | |
5e745b04 AK |
3337 | struct ext4_prealloc_space *pa, *cpa = NULL; |
3338 | ext4_fsblk_t goal_block; | |
c9de560d AT |
3339 | |
3340 | /* only data can be preallocated */ | |
3341 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
3342 | return 0; | |
3343 | ||
3344 | /* first, try per-file preallocation */ | |
3345 | rcu_read_lock(); | |
9a0762c5 | 3346 | list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { |
c9de560d AT |
3347 | |
3348 | /* all fields in this condition don't change, | |
3349 | * so we can skip locking for them */ | |
3350 | if (ac->ac_o_ex.fe_logical < pa->pa_lstart || | |
53accfa9 TT |
3351 | ac->ac_o_ex.fe_logical >= (pa->pa_lstart + |
3352 | EXT4_C2B(sbi, pa->pa_len))) | |
c9de560d AT |
3353 | continue; |
3354 | ||
fb0a387d | 3355 | /* non-extent files can't have physical blocks past 2^32 */ |
12e9b892 | 3356 | if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) && |
53accfa9 TT |
3357 | (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) > |
3358 | EXT4_MAX_BLOCK_FILE_PHYS)) | |
fb0a387d ES |
3359 | continue; |
3360 | ||
c9de560d AT |
3361 | /* found preallocated blocks, use them */ |
3362 | spin_lock(&pa->pa_lock); | |
3363 | if (pa->pa_deleted == 0 && pa->pa_free) { | |
3364 | atomic_inc(&pa->pa_count); | |
3365 | ext4_mb_use_inode_pa(ac, pa); | |
3366 | spin_unlock(&pa->pa_lock); | |
3367 | ac->ac_criteria = 10; | |
3368 | rcu_read_unlock(); | |
3369 | return 1; | |
3370 | } | |
3371 | spin_unlock(&pa->pa_lock); | |
3372 | } | |
3373 | rcu_read_unlock(); | |
3374 | ||
3375 | /* can we use group allocation? */ | |
3376 | if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)) | |
3377 | return 0; | |
3378 | ||
3379 | /* inode may have no locality group for some reason */ | |
3380 | lg = ac->ac_lg; | |
3381 | if (lg == NULL) | |
3382 | return 0; | |
6be2ded1 AK |
3383 | order = fls(ac->ac_o_ex.fe_len) - 1; |
3384 | if (order > PREALLOC_TB_SIZE - 1) | |
3385 | /* The max size of hash table is PREALLOC_TB_SIZE */ | |
3386 | order = PREALLOC_TB_SIZE - 1; | |
3387 | ||
bda00de7 | 3388 | goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex); |
5e745b04 AK |
3389 | /* |
3390 | * search for the prealloc space that is having | |
3391 | * minimal distance from the goal block. | |
3392 | */ | |
6be2ded1 AK |
3393 | for (i = order; i < PREALLOC_TB_SIZE; i++) { |
3394 | rcu_read_lock(); | |
3395 | list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i], | |
3396 | pa_inode_list) { | |
3397 | spin_lock(&pa->pa_lock); | |
3398 | if (pa->pa_deleted == 0 && | |
3399 | pa->pa_free >= ac->ac_o_ex.fe_len) { | |
5e745b04 AK |
3400 | |
3401 | cpa = ext4_mb_check_group_pa(goal_block, | |
3402 | pa, cpa); | |
6be2ded1 | 3403 | } |
c9de560d | 3404 | spin_unlock(&pa->pa_lock); |
c9de560d | 3405 | } |
6be2ded1 | 3406 | rcu_read_unlock(); |
c9de560d | 3407 | } |
5e745b04 AK |
3408 | if (cpa) { |
3409 | ext4_mb_use_group_pa(ac, cpa); | |
3410 | ac->ac_criteria = 20; | |
3411 | return 1; | |
3412 | } | |
c9de560d AT |
3413 | return 0; |
3414 | } | |
3415 | ||
7a2fcbf7 AK |
3416 | /* |
3417 | * the function goes through all block freed in the group | |
3418 | * but not yet committed and marks them used in in-core bitmap. | |
3419 | * buddy must be generated from this bitmap | |
955ce5f5 | 3420 | * Need to be called with the ext4 group lock held |
7a2fcbf7 AK |
3421 | */ |
3422 | static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, | |
3423 | ext4_group_t group) | |
3424 | { | |
3425 | struct rb_node *n; | |
3426 | struct ext4_group_info *grp; | |
3427 | struct ext4_free_data *entry; | |
3428 | ||
3429 | grp = ext4_get_group_info(sb, group); | |
3430 | n = rb_first(&(grp->bb_free_root)); | |
3431 | ||
3432 | while (n) { | |
18aadd47 BJ |
3433 | entry = rb_entry(n, struct ext4_free_data, efd_node); |
3434 | ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count); | |
7a2fcbf7 AK |
3435 | n = rb_next(n); |
3436 | } | |
3437 | return; | |
3438 | } | |
3439 | ||
c9de560d AT |
3440 | /* |
3441 | * the function goes through all preallocation in this group and marks them | |
3442 | * used in in-core bitmap. buddy must be generated from this bitmap | |
955ce5f5 | 3443 | * Need to be called with ext4 group lock held |
c9de560d | 3444 | */ |
089ceecc ES |
3445 | static noinline_for_stack |
3446 | void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, | |
c9de560d AT |
3447 | ext4_group_t group) |
3448 | { | |
3449 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
3450 | struct ext4_prealloc_space *pa; | |
3451 | struct list_head *cur; | |
3452 | ext4_group_t groupnr; | |
3453 | ext4_grpblk_t start; | |
3454 | int preallocated = 0; | |
c9de560d AT |
3455 | int len; |
3456 | ||
3457 | /* all form of preallocation discards first load group, | |
3458 | * so the only competing code is preallocation use. | |
3459 | * we don't need any locking here | |
3460 | * notice we do NOT ignore preallocations with pa_deleted | |
3461 | * otherwise we could leave used blocks available for | |
3462 | * allocation in buddy when concurrent ext4_mb_put_pa() | |
3463 | * is dropping preallocation | |
3464 | */ | |
3465 | list_for_each(cur, &grp->bb_prealloc_list) { | |
3466 | pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); | |
3467 | spin_lock(&pa->pa_lock); | |
3468 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, | |
3469 | &groupnr, &start); | |
3470 | len = pa->pa_len; | |
3471 | spin_unlock(&pa->pa_lock); | |
3472 | if (unlikely(len == 0)) | |
3473 | continue; | |
3474 | BUG_ON(groupnr != group); | |
c3e94d1d | 3475 | ext4_set_bits(bitmap, start, len); |
c9de560d | 3476 | preallocated += len; |
c9de560d | 3477 | } |
6ba495e9 | 3478 | mb_debug(1, "prellocated %u for group %u\n", preallocated, group); |
c9de560d AT |
3479 | } |
3480 | ||
3481 | static void ext4_mb_pa_callback(struct rcu_head *head) | |
3482 | { | |
3483 | struct ext4_prealloc_space *pa; | |
3484 | pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu); | |
4e8d2139 JR |
3485 | |
3486 | BUG_ON(atomic_read(&pa->pa_count)); | |
3487 | BUG_ON(pa->pa_deleted == 0); | |
c9de560d AT |
3488 | kmem_cache_free(ext4_pspace_cachep, pa); |
3489 | } | |
3490 | ||
3491 | /* | |
3492 | * drops a reference to preallocated space descriptor | |
3493 | * if this was the last reference and the space is consumed | |
3494 | */ | |
3495 | static void ext4_mb_put_pa(struct ext4_allocation_context *ac, | |
3496 | struct super_block *sb, struct ext4_prealloc_space *pa) | |
3497 | { | |
a9df9a49 | 3498 | ext4_group_t grp; |
d33a1976 | 3499 | ext4_fsblk_t grp_blk; |
c9de560d | 3500 | |
c9de560d AT |
3501 | /* in this short window concurrent discard can set pa_deleted */ |
3502 | spin_lock(&pa->pa_lock); | |
4e8d2139 JR |
3503 | if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) { |
3504 | spin_unlock(&pa->pa_lock); | |
3505 | return; | |
3506 | } | |
3507 | ||
c9de560d AT |
3508 | if (pa->pa_deleted == 1) { |
3509 | spin_unlock(&pa->pa_lock); | |
3510 | return; | |
3511 | } | |
3512 | ||
3513 | pa->pa_deleted = 1; | |
3514 | spin_unlock(&pa->pa_lock); | |
3515 | ||
d33a1976 | 3516 | grp_blk = pa->pa_pstart; |
60e6679e | 3517 | /* |
cc0fb9ad AK |
3518 | * If doing group-based preallocation, pa_pstart may be in the |
3519 | * next group when pa is used up | |
3520 | */ | |
3521 | if (pa->pa_type == MB_GROUP_PA) | |
d33a1976 ES |
3522 | grp_blk--; |
3523 | ||
bd86298e | 3524 | grp = ext4_get_group_number(sb, grp_blk); |
c9de560d AT |
3525 | |
3526 | /* | |
3527 | * possible race: | |
3528 | * | |
3529 | * P1 (buddy init) P2 (regular allocation) | |
3530 | * find block B in PA | |
3531 | * copy on-disk bitmap to buddy | |
3532 | * mark B in on-disk bitmap | |
3533 | * drop PA from group | |
3534 | * mark all PAs in buddy | |
3535 | * | |
3536 | * thus, P1 initializes buddy with B available. to prevent this | |
3537 | * we make "copy" and "mark all PAs" atomic and serialize "drop PA" | |
3538 | * against that pair | |
3539 | */ | |
3540 | ext4_lock_group(sb, grp); | |
3541 | list_del(&pa->pa_group_list); | |
3542 | ext4_unlock_group(sb, grp); | |
3543 | ||
3544 | spin_lock(pa->pa_obj_lock); | |
3545 | list_del_rcu(&pa->pa_inode_list); | |
3546 | spin_unlock(pa->pa_obj_lock); | |
3547 | ||
3548 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
3549 | } | |
3550 | ||
3551 | /* | |
3552 | * creates new preallocated space for given inode | |
3553 | */ | |
4ddfef7b ES |
3554 | static noinline_for_stack int |
3555 | ext4_mb_new_inode_pa(struct ext4_allocation_context *ac) | |
c9de560d AT |
3556 | { |
3557 | struct super_block *sb = ac->ac_sb; | |
53accfa9 | 3558 | struct ext4_sb_info *sbi = EXT4_SB(sb); |
c9de560d AT |
3559 | struct ext4_prealloc_space *pa; |
3560 | struct ext4_group_info *grp; | |
3561 | struct ext4_inode_info *ei; | |
3562 | ||
3563 | /* preallocate only when found space is larger then requested */ | |
3564 | BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); | |
3565 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
3566 | BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); | |
3567 | ||
3568 | pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS); | |
3569 | if (pa == NULL) | |
3570 | return -ENOMEM; | |
3571 | ||
3572 | if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) { | |
3573 | int winl; | |
3574 | int wins; | |
3575 | int win; | |
3576 | int offs; | |
3577 | ||
3578 | /* we can't allocate as much as normalizer wants. | |
3579 | * so, found space must get proper lstart | |
3580 | * to cover original request */ | |
3581 | BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical); | |
3582 | BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len); | |
3583 | ||
3584 | /* we're limited by original request in that | |
3585 | * logical block must be covered any way | |
3586 | * winl is window we can move our chunk within */ | |
3587 | winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical; | |
3588 | ||
3589 | /* also, we should cover whole original request */ | |
53accfa9 | 3590 | wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len); |
c9de560d AT |
3591 | |
3592 | /* the smallest one defines real window */ | |
3593 | win = min(winl, wins); | |
3594 | ||
53accfa9 TT |
3595 | offs = ac->ac_o_ex.fe_logical % |
3596 | EXT4_C2B(sbi, ac->ac_b_ex.fe_len); | |
c9de560d AT |
3597 | if (offs && offs < win) |
3598 | win = offs; | |
3599 | ||
53accfa9 | 3600 | ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - |
810da240 | 3601 | EXT4_NUM_B2C(sbi, win); |
c9de560d AT |
3602 | BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical); |
3603 | BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len); | |
3604 | } | |
3605 | ||
3606 | /* preallocation can change ac_b_ex, thus we store actually | |
3607 | * allocated blocks for history */ | |
3608 | ac->ac_f_ex = ac->ac_b_ex; | |
3609 | ||
3610 | pa->pa_lstart = ac->ac_b_ex.fe_logical; | |
3611 | pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); | |
3612 | pa->pa_len = ac->ac_b_ex.fe_len; | |
3613 | pa->pa_free = pa->pa_len; | |
3614 | atomic_set(&pa->pa_count, 1); | |
3615 | spin_lock_init(&pa->pa_lock); | |
d794bf8e AK |
3616 | INIT_LIST_HEAD(&pa->pa_inode_list); |
3617 | INIT_LIST_HEAD(&pa->pa_group_list); | |
c9de560d | 3618 | pa->pa_deleted = 0; |
cc0fb9ad | 3619 | pa->pa_type = MB_INODE_PA; |
c9de560d | 3620 | |
6ba495e9 | 3621 | mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa, |
c9de560d | 3622 | pa->pa_pstart, pa->pa_len, pa->pa_lstart); |
9bffad1e | 3623 | trace_ext4_mb_new_inode_pa(ac, pa); |
c9de560d AT |
3624 | |
3625 | ext4_mb_use_inode_pa(ac, pa); | |
53accfa9 | 3626 | atomic_add(pa->pa_free, &sbi->s_mb_preallocated); |
c9de560d AT |
3627 | |
3628 | ei = EXT4_I(ac->ac_inode); | |
3629 | grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); | |
3630 | ||
3631 | pa->pa_obj_lock = &ei->i_prealloc_lock; | |
3632 | pa->pa_inode = ac->ac_inode; | |
3633 | ||
3634 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
3635 | list_add(&pa->pa_group_list, &grp->bb_prealloc_list); | |
3636 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
3637 | ||
3638 | spin_lock(pa->pa_obj_lock); | |
3639 | list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list); | |
3640 | spin_unlock(pa->pa_obj_lock); | |
3641 | ||
3642 | return 0; | |
3643 | } | |
3644 | ||
3645 | /* | |
3646 | * creates new preallocated space for locality group inodes belongs to | |
3647 | */ | |
4ddfef7b ES |
3648 | static noinline_for_stack int |
3649 | ext4_mb_new_group_pa(struct ext4_allocation_context *ac) | |
c9de560d AT |
3650 | { |
3651 | struct super_block *sb = ac->ac_sb; | |
3652 | struct ext4_locality_group *lg; | |
3653 | struct ext4_prealloc_space *pa; | |
3654 | struct ext4_group_info *grp; | |
3655 | ||
3656 | /* preallocate only when found space is larger then requested */ | |
3657 | BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); | |
3658 | BUG_ON(ac->ac_status != AC_STATUS_FOUND); | |
3659 | BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); | |
3660 | ||
3661 | BUG_ON(ext4_pspace_cachep == NULL); | |
3662 | pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS); | |
3663 | if (pa == NULL) | |
3664 | return -ENOMEM; | |
3665 | ||
3666 | /* preallocation can change ac_b_ex, thus we store actually | |
3667 | * allocated blocks for history */ | |
3668 | ac->ac_f_ex = ac->ac_b_ex; | |
3669 | ||
3670 | pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); | |
3671 | pa->pa_lstart = pa->pa_pstart; | |
3672 | pa->pa_len = ac->ac_b_ex.fe_len; | |
3673 | pa->pa_free = pa->pa_len; | |
3674 | atomic_set(&pa->pa_count, 1); | |
3675 | spin_lock_init(&pa->pa_lock); | |
6be2ded1 | 3676 | INIT_LIST_HEAD(&pa->pa_inode_list); |
d794bf8e | 3677 | INIT_LIST_HEAD(&pa->pa_group_list); |
c9de560d | 3678 | pa->pa_deleted = 0; |
cc0fb9ad | 3679 | pa->pa_type = MB_GROUP_PA; |
c9de560d | 3680 | |
6ba495e9 | 3681 | mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa, |
9bffad1e TT |
3682 | pa->pa_pstart, pa->pa_len, pa->pa_lstart); |
3683 | trace_ext4_mb_new_group_pa(ac, pa); | |
c9de560d AT |
3684 | |
3685 | ext4_mb_use_group_pa(ac, pa); | |
3686 | atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated); | |
3687 | ||
3688 | grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); | |
3689 | lg = ac->ac_lg; | |
3690 | BUG_ON(lg == NULL); | |
3691 | ||
3692 | pa->pa_obj_lock = &lg->lg_prealloc_lock; | |
3693 | pa->pa_inode = NULL; | |
3694 | ||
3695 | ext4_lock_group(sb, ac->ac_b_ex.fe_group); | |
3696 | list_add(&pa->pa_group_list, &grp->bb_prealloc_list); | |
3697 | ext4_unlock_group(sb, ac->ac_b_ex.fe_group); | |
3698 | ||
6be2ded1 AK |
3699 | /* |
3700 | * We will later add the new pa to the right bucket | |
3701 | * after updating the pa_free in ext4_mb_release_context | |
3702 | */ | |
c9de560d AT |
3703 | return 0; |
3704 | } | |
3705 | ||
3706 | static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac) | |
3707 | { | |
3708 | int err; | |
3709 | ||
3710 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) | |
3711 | err = ext4_mb_new_group_pa(ac); | |
3712 | else | |
3713 | err = ext4_mb_new_inode_pa(ac); | |
3714 | return err; | |
3715 | } | |
3716 | ||
3717 | /* | |
3718 | * finds all unused blocks in on-disk bitmap, frees them in | |
3719 | * in-core bitmap and buddy. | |
3720 | * @pa must be unlinked from inode and group lists, so that | |
3721 | * nobody else can find/use it. | |
3722 | * the caller MUST hold group/inode locks. | |
3723 | * TODO: optimize the case when there are no in-core structures yet | |
3724 | */ | |
4ddfef7b ES |
3725 | static noinline_for_stack int |
3726 | ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh, | |
3e1e5f50 | 3727 | struct ext4_prealloc_space *pa) |
c9de560d | 3728 | { |
c9de560d AT |
3729 | struct super_block *sb = e4b->bd_sb; |
3730 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
498e5f24 TT |
3731 | unsigned int end; |
3732 | unsigned int next; | |
c9de560d AT |
3733 | ext4_group_t group; |
3734 | ext4_grpblk_t bit; | |
ba80b101 | 3735 | unsigned long long grp_blk_start; |
c9de560d AT |
3736 | int err = 0; |
3737 | int free = 0; | |
3738 | ||
3739 | BUG_ON(pa->pa_deleted == 0); | |
3740 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); | |
53accfa9 | 3741 | grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit); |
c9de560d AT |
3742 | BUG_ON(group != e4b->bd_group && pa->pa_len != 0); |
3743 | end = bit + pa->pa_len; | |
3744 | ||
c9de560d | 3745 | while (bit < end) { |
ffad0a44 | 3746 | bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit); |
c9de560d AT |
3747 | if (bit >= end) |
3748 | break; | |
ffad0a44 | 3749 | next = mb_find_next_bit(bitmap_bh->b_data, end, bit); |
6ba495e9 | 3750 | mb_debug(1, " free preallocated %u/%u in group %u\n", |
5a0790c2 AK |
3751 | (unsigned) ext4_group_first_block_no(sb, group) + bit, |
3752 | (unsigned) next - bit, (unsigned) group); | |
c9de560d AT |
3753 | free += next - bit; |
3754 | ||
3e1e5f50 | 3755 | trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit); |
53accfa9 TT |
3756 | trace_ext4_mb_release_inode_pa(pa, (grp_blk_start + |
3757 | EXT4_C2B(sbi, bit)), | |
a9c667f8 | 3758 | next - bit); |
c9de560d AT |
3759 | mb_free_blocks(pa->pa_inode, e4b, bit, next - bit); |
3760 | bit = next + 1; | |
3761 | } | |
3762 | if (free != pa->pa_free) { | |
9d8b9ec4 TT |
3763 | ext4_msg(e4b->bd_sb, KERN_CRIT, |
3764 | "pa %p: logic %lu, phys. %lu, len %lu", | |
3765 | pa, (unsigned long) pa->pa_lstart, | |
3766 | (unsigned long) pa->pa_pstart, | |
3767 | (unsigned long) pa->pa_len); | |
e29136f8 | 3768 | ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u", |
5d1b1b3f | 3769 | free, pa->pa_free); |
e56eb659 AK |
3770 | /* |
3771 | * pa is already deleted so we use the value obtained | |
3772 | * from the bitmap and continue. | |
3773 | */ | |
c9de560d | 3774 | } |
c9de560d AT |
3775 | atomic_add(free, &sbi->s_mb_discarded); |
3776 | ||
3777 | return err; | |
3778 | } | |
3779 | ||
4ddfef7b ES |
3780 | static noinline_for_stack int |
3781 | ext4_mb_release_group_pa(struct ext4_buddy *e4b, | |
3e1e5f50 | 3782 | struct ext4_prealloc_space *pa) |
c9de560d | 3783 | { |
c9de560d AT |
3784 | struct super_block *sb = e4b->bd_sb; |
3785 | ext4_group_t group; | |
3786 | ext4_grpblk_t bit; | |
3787 | ||
60e07cf5 | 3788 | trace_ext4_mb_release_group_pa(sb, pa); |
c9de560d AT |
3789 | BUG_ON(pa->pa_deleted == 0); |
3790 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); | |
3791 | BUG_ON(group != e4b->bd_group && pa->pa_len != 0); | |
3792 | mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len); | |
3793 | atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded); | |
3e1e5f50 | 3794 | trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len); |
c9de560d AT |
3795 | |
3796 | return 0; | |
3797 | } | |
3798 | ||
3799 | /* | |
3800 | * releases all preallocations in given group | |
3801 | * | |
3802 | * first, we need to decide discard policy: | |
3803 | * - when do we discard | |
3804 | * 1) ENOSPC | |
3805 | * - how many do we discard | |
3806 | * 1) how many requested | |
3807 | */ | |
4ddfef7b ES |
3808 | static noinline_for_stack int |
3809 | ext4_mb_discard_group_preallocations(struct super_block *sb, | |
c9de560d AT |
3810 | ext4_group_t group, int needed) |
3811 | { | |
3812 | struct ext4_group_info *grp = ext4_get_group_info(sb, group); | |
3813 | struct buffer_head *bitmap_bh = NULL; | |
3814 | struct ext4_prealloc_space *pa, *tmp; | |
3815 | struct list_head list; | |
3816 | struct ext4_buddy e4b; | |
3817 | int err; | |
3818 | int busy = 0; | |
3819 | int free = 0; | |
3820 | ||
6ba495e9 | 3821 | mb_debug(1, "discard preallocation for group %u\n", group); |
c9de560d AT |
3822 | |
3823 | if (list_empty(&grp->bb_prealloc_list)) | |
3824 | return 0; | |
3825 | ||
574ca174 | 3826 | bitmap_bh = ext4_read_block_bitmap(sb, group); |
c9de560d | 3827 | if (bitmap_bh == NULL) { |
12062ddd | 3828 | ext4_error(sb, "Error reading block bitmap for %u", group); |
ce89f46c | 3829 | return 0; |
c9de560d AT |
3830 | } |
3831 | ||
3832 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
ce89f46c | 3833 | if (err) { |
12062ddd | 3834 | ext4_error(sb, "Error loading buddy information for %u", group); |
ce89f46c AK |
3835 | put_bh(bitmap_bh); |
3836 | return 0; | |
3837 | } | |
c9de560d AT |
3838 | |
3839 | if (needed == 0) | |
7137d7a4 | 3840 | needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1; |
c9de560d | 3841 | |
c9de560d | 3842 | INIT_LIST_HEAD(&list); |
c9de560d AT |
3843 | repeat: |
3844 | ext4_lock_group(sb, group); | |
3845 | list_for_each_entry_safe(pa, tmp, | |
3846 | &grp->bb_prealloc_list, pa_group_list) { | |
3847 | spin_lock(&pa->pa_lock); | |
3848 | if (atomic_read(&pa->pa_count)) { | |
3849 | spin_unlock(&pa->pa_lock); | |
3850 | busy = 1; | |
3851 | continue; | |
3852 | } | |
3853 | if (pa->pa_deleted) { | |
3854 | spin_unlock(&pa->pa_lock); | |
3855 | continue; | |
3856 | } | |
3857 | ||
3858 | /* seems this one can be freed ... */ | |
3859 | pa->pa_deleted = 1; | |
3860 | ||
3861 | /* we can trust pa_free ... */ | |
3862 | free += pa->pa_free; | |
3863 | ||
3864 | spin_unlock(&pa->pa_lock); | |
3865 | ||
3866 | list_del(&pa->pa_group_list); | |
3867 | list_add(&pa->u.pa_tmp_list, &list); | |
3868 | } | |
3869 | ||
3870 | /* if we still need more blocks and some PAs were used, try again */ | |
3871 | if (free < needed && busy) { | |
3872 | busy = 0; | |
3873 | ext4_unlock_group(sb, group); | |
bb8b20ed | 3874 | cond_resched(); |
c9de560d AT |
3875 | goto repeat; |
3876 | } | |
3877 | ||
3878 | /* found anything to free? */ | |
3879 | if (list_empty(&list)) { | |
3880 | BUG_ON(free != 0); | |
3881 | goto out; | |
3882 | } | |
3883 | ||
3884 | /* now free all selected PAs */ | |
3885 | list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { | |
3886 | ||
3887 | /* remove from object (inode or locality group) */ | |
3888 | spin_lock(pa->pa_obj_lock); | |
3889 | list_del_rcu(&pa->pa_inode_list); | |
3890 | spin_unlock(pa->pa_obj_lock); | |
3891 | ||
cc0fb9ad | 3892 | if (pa->pa_type == MB_GROUP_PA) |
3e1e5f50 | 3893 | ext4_mb_release_group_pa(&e4b, pa); |
c9de560d | 3894 | else |
3e1e5f50 | 3895 | ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); |
c9de560d AT |
3896 | |
3897 | list_del(&pa->u.pa_tmp_list); | |
3898 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
3899 | } | |
3900 | ||
3901 | out: | |
3902 | ext4_unlock_group(sb, group); | |
e39e07fd | 3903 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
3904 | put_bh(bitmap_bh); |
3905 | return free; | |
3906 | } | |
3907 | ||
3908 | /* | |
3909 | * releases all non-used preallocated blocks for given inode | |
3910 | * | |
3911 | * It's important to discard preallocations under i_data_sem | |
3912 | * We don't want another block to be served from the prealloc | |
3913 | * space when we are discarding the inode prealloc space. | |
3914 | * | |
3915 | * FIXME!! Make sure it is valid at all the call sites | |
3916 | */ | |
c2ea3fde | 3917 | void ext4_discard_preallocations(struct inode *inode) |
c9de560d AT |
3918 | { |
3919 | struct ext4_inode_info *ei = EXT4_I(inode); | |
3920 | struct super_block *sb = inode->i_sb; | |
3921 | struct buffer_head *bitmap_bh = NULL; | |
3922 | struct ext4_prealloc_space *pa, *tmp; | |
3923 | ext4_group_t group = 0; | |
3924 | struct list_head list; | |
3925 | struct ext4_buddy e4b; | |
3926 | int err; | |
3927 | ||
c2ea3fde | 3928 | if (!S_ISREG(inode->i_mode)) { |
c9de560d AT |
3929 | /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/ |
3930 | return; | |
3931 | } | |
3932 | ||
6ba495e9 | 3933 | mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino); |
9bffad1e | 3934 | trace_ext4_discard_preallocations(inode); |
c9de560d AT |
3935 | |
3936 | INIT_LIST_HEAD(&list); | |
3937 | ||
3938 | repeat: | |
3939 | /* first, collect all pa's in the inode */ | |
3940 | spin_lock(&ei->i_prealloc_lock); | |
3941 | while (!list_empty(&ei->i_prealloc_list)) { | |
3942 | pa = list_entry(ei->i_prealloc_list.next, | |
3943 | struct ext4_prealloc_space, pa_inode_list); | |
3944 | BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock); | |
3945 | spin_lock(&pa->pa_lock); | |
3946 | if (atomic_read(&pa->pa_count)) { | |
3947 | /* this shouldn't happen often - nobody should | |
3948 | * use preallocation while we're discarding it */ | |
3949 | spin_unlock(&pa->pa_lock); | |
3950 | spin_unlock(&ei->i_prealloc_lock); | |
9d8b9ec4 TT |
3951 | ext4_msg(sb, KERN_ERR, |
3952 | "uh-oh! used pa while discarding"); | |
c9de560d AT |
3953 | WARN_ON(1); |
3954 | schedule_timeout_uninterruptible(HZ); | |
3955 | goto repeat; | |
3956 | ||
3957 | } | |
3958 | if (pa->pa_deleted == 0) { | |
3959 | pa->pa_deleted = 1; | |
3960 | spin_unlock(&pa->pa_lock); | |
3961 | list_del_rcu(&pa->pa_inode_list); | |
3962 | list_add(&pa->u.pa_tmp_list, &list); | |
3963 | continue; | |
3964 | } | |
3965 | ||
3966 | /* someone is deleting pa right now */ | |
3967 | spin_unlock(&pa->pa_lock); | |
3968 | spin_unlock(&ei->i_prealloc_lock); | |
3969 | ||
3970 | /* we have to wait here because pa_deleted | |
3971 | * doesn't mean pa is already unlinked from | |
3972 | * the list. as we might be called from | |
3973 | * ->clear_inode() the inode will get freed | |
3974 | * and concurrent thread which is unlinking | |
3975 | * pa from inode's list may access already | |
3976 | * freed memory, bad-bad-bad */ | |
3977 | ||
3978 | /* XXX: if this happens too often, we can | |
3979 | * add a flag to force wait only in case | |
3980 | * of ->clear_inode(), but not in case of | |
3981 | * regular truncate */ | |
3982 | schedule_timeout_uninterruptible(HZ); | |
3983 | goto repeat; | |
3984 | } | |
3985 | spin_unlock(&ei->i_prealloc_lock); | |
3986 | ||
3987 | list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { | |
cc0fb9ad | 3988 | BUG_ON(pa->pa_type != MB_INODE_PA); |
bd86298e | 3989 | group = ext4_get_group_number(sb, pa->pa_pstart); |
c9de560d AT |
3990 | |
3991 | err = ext4_mb_load_buddy(sb, group, &e4b); | |
ce89f46c | 3992 | if (err) { |
12062ddd ES |
3993 | ext4_error(sb, "Error loading buddy information for %u", |
3994 | group); | |
ce89f46c AK |
3995 | continue; |
3996 | } | |
c9de560d | 3997 | |
574ca174 | 3998 | bitmap_bh = ext4_read_block_bitmap(sb, group); |
c9de560d | 3999 | if (bitmap_bh == NULL) { |
12062ddd ES |
4000 | ext4_error(sb, "Error reading block bitmap for %u", |
4001 | group); | |
e39e07fd | 4002 | ext4_mb_unload_buddy(&e4b); |
ce89f46c | 4003 | continue; |
c9de560d AT |
4004 | } |
4005 | ||
4006 | ext4_lock_group(sb, group); | |
4007 | list_del(&pa->pa_group_list); | |
3e1e5f50 | 4008 | ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); |
c9de560d AT |
4009 | ext4_unlock_group(sb, group); |
4010 | ||
e39e07fd | 4011 | ext4_mb_unload_buddy(&e4b); |
c9de560d AT |
4012 | put_bh(bitmap_bh); |
4013 | ||
4014 | list_del(&pa->u.pa_tmp_list); | |
4015 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
4016 | } | |
4017 | } | |
4018 | ||
6ba495e9 | 4019 | #ifdef CONFIG_EXT4_DEBUG |
c9de560d AT |
4020 | static void ext4_mb_show_ac(struct ext4_allocation_context *ac) |
4021 | { | |
4022 | struct super_block *sb = ac->ac_sb; | |
8df9675f | 4023 | ext4_group_t ngroups, i; |
c9de560d | 4024 | |
a0b30c12 | 4025 | if (!ext4_mballoc_debug || |
4dd89fc6 | 4026 | (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)) |
e3570639 ES |
4027 | return; |
4028 | ||
7f6a11e7 | 4029 | ext4_msg(ac->ac_sb, KERN_ERR, "Can't allocate:" |
9d8b9ec4 | 4030 | " Allocation context details:"); |
7f6a11e7 | 4031 | ext4_msg(ac->ac_sb, KERN_ERR, "status %d flags %d", |
c9de560d | 4032 | ac->ac_status, ac->ac_flags); |
7f6a11e7 | 4033 | ext4_msg(ac->ac_sb, KERN_ERR, "orig %lu/%lu/%lu@%lu, " |
9d8b9ec4 TT |
4034 | "goal %lu/%lu/%lu@%lu, " |
4035 | "best %lu/%lu/%lu@%lu cr %d", | |
c9de560d AT |
4036 | (unsigned long)ac->ac_o_ex.fe_group, |
4037 | (unsigned long)ac->ac_o_ex.fe_start, | |
4038 | (unsigned long)ac->ac_o_ex.fe_len, | |
4039 | (unsigned long)ac->ac_o_ex.fe_logical, | |
4040 | (unsigned long)ac->ac_g_ex.fe_group, | |
4041 | (unsigned long)ac->ac_g_ex.fe_start, | |
4042 | (unsigned long)ac->ac_g_ex.fe_len, | |
4043 | (unsigned long)ac->ac_g_ex.fe_logical, | |
4044 | (unsigned long)ac->ac_b_ex.fe_group, | |
4045 | (unsigned long)ac->ac_b_ex.fe_start, | |
4046 | (unsigned long)ac->ac_b_ex.fe_len, | |
4047 | (unsigned long)ac->ac_b_ex.fe_logical, | |
4048 | (int)ac->ac_criteria); | |
dc9ddd98 | 4049 | ext4_msg(ac->ac_sb, KERN_ERR, "%d found", ac->ac_found); |
7f6a11e7 | 4050 | ext4_msg(ac->ac_sb, KERN_ERR, "groups: "); |
8df9675f TT |
4051 | ngroups = ext4_get_groups_count(sb); |
4052 | for (i = 0; i < ngroups; i++) { | |
c9de560d AT |
4053 | struct ext4_group_info *grp = ext4_get_group_info(sb, i); |
4054 | struct ext4_prealloc_space *pa; | |
4055 | ext4_grpblk_t start; | |
4056 | struct list_head *cur; | |
4057 | ext4_lock_group(sb, i); | |
4058 | list_for_each(cur, &grp->bb_prealloc_list) { | |
4059 | pa = list_entry(cur, struct ext4_prealloc_space, | |
4060 | pa_group_list); | |
4061 | spin_lock(&pa->pa_lock); | |
4062 | ext4_get_group_no_and_offset(sb, pa->pa_pstart, | |
4063 | NULL, &start); | |
4064 | spin_unlock(&pa->pa_lock); | |
1c718505 AF |
4065 | printk(KERN_ERR "PA:%u:%d:%u \n", i, |
4066 | start, pa->pa_len); | |
c9de560d | 4067 | } |
60bd63d1 | 4068 | ext4_unlock_group(sb, i); |
c9de560d AT |
4069 | |
4070 | if (grp->bb_free == 0) | |
4071 | continue; | |
1c718505 | 4072 | printk(KERN_ERR "%u: %d/%d \n", |
c9de560d AT |
4073 | i, grp->bb_free, grp->bb_fragments); |
4074 | } | |
4075 | printk(KERN_ERR "\n"); | |
4076 | } | |
4077 | #else | |
4078 | static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac) | |
4079 | { | |
4080 | return; | |
4081 | } | |
4082 | #endif | |
4083 | ||
4084 | /* | |
4085 | * We use locality group preallocation for small size file. The size of the | |
4086 | * file is determined by the current size or the resulting size after | |
4087 | * allocation which ever is larger | |
4088 | * | |
b713a5ec | 4089 | * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req |
c9de560d AT |
4090 | */ |
4091 | static void ext4_mb_group_or_file(struct ext4_allocation_context *ac) | |
4092 | { | |
4093 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); | |
4094 | int bsbits = ac->ac_sb->s_blocksize_bits; | |
4095 | loff_t size, isize; | |
4096 | ||
4097 | if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) | |
4098 | return; | |
4099 | ||
4ba74d00 TT |
4100 | if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) |
4101 | return; | |
4102 | ||
53accfa9 | 4103 | size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len); |
50797481 TT |
4104 | isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1) |
4105 | >> bsbits; | |
c9de560d | 4106 | |
50797481 TT |
4107 | if ((size == isize) && |
4108 | !ext4_fs_is_busy(sbi) && | |
4109 | (atomic_read(&ac->ac_inode->i_writecount) == 0)) { | |
4110 | ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC; | |
4111 | return; | |
4112 | } | |
4113 | ||
ebbe0277 RD |
4114 | if (sbi->s_mb_group_prealloc <= 0) { |
4115 | ac->ac_flags |= EXT4_MB_STREAM_ALLOC; | |
4116 | return; | |
4117 | } | |
4118 | ||
c9de560d | 4119 | /* don't use group allocation for large files */ |
71780577 | 4120 | size = max(size, isize); |
cc483f10 | 4121 | if (size > sbi->s_mb_stream_request) { |
4ba74d00 | 4122 | ac->ac_flags |= EXT4_MB_STREAM_ALLOC; |
c9de560d | 4123 | return; |
4ba74d00 | 4124 | } |
c9de560d AT |
4125 | |
4126 | BUG_ON(ac->ac_lg != NULL); | |
4127 | /* | |
4128 | * locality group prealloc space are per cpu. The reason for having | |
4129 | * per cpu locality group is to reduce the contention between block | |
4130 | * request from multiple CPUs. | |
4131 | */ | |
ca0c9584 | 4132 | ac->ac_lg = __this_cpu_ptr(sbi->s_locality_groups); |
c9de560d AT |
4133 | |
4134 | /* we're going to use group allocation */ | |
4135 | ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC; | |
4136 | ||
4137 | /* serialize all allocations in the group */ | |
4138 | mutex_lock(&ac->ac_lg->lg_mutex); | |
4139 | } | |
4140 | ||
4ddfef7b ES |
4141 | static noinline_for_stack int |
4142 | ext4_mb_initialize_context(struct ext4_allocation_context *ac, | |
c9de560d AT |
4143 | struct ext4_allocation_request *ar) |
4144 | { | |
4145 | struct super_block *sb = ar->inode->i_sb; | |
4146 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
4147 | struct ext4_super_block *es = sbi->s_es; | |
4148 | ext4_group_t group; | |
498e5f24 TT |
4149 | unsigned int len; |
4150 | ext4_fsblk_t goal; | |
c9de560d AT |
4151 | ext4_grpblk_t block; |
4152 | ||
4153 | /* we can't allocate > group size */ | |
4154 | len = ar->len; | |
4155 | ||
4156 | /* just a dirty hack to filter too big requests */ | |
40ae3487 TT |
4157 | if (len >= EXT4_CLUSTERS_PER_GROUP(sb)) |
4158 | len = EXT4_CLUSTERS_PER_GROUP(sb); | |
c9de560d AT |
4159 | |
4160 | /* start searching from the goal */ | |
4161 | goal = ar->goal; | |
4162 | if (goal < le32_to_cpu(es->s_first_data_block) || | |
4163 | goal >= ext4_blocks_count(es)) | |
4164 | goal = le32_to_cpu(es->s_first_data_block); | |
4165 | ext4_get_group_no_and_offset(sb, goal, &group, &block); | |
4166 | ||
4167 | /* set up allocation goals */ | |
f5a44db5 | 4168 | ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical); |
c9de560d | 4169 | ac->ac_status = AC_STATUS_CONTINUE; |
c9de560d AT |
4170 | ac->ac_sb = sb; |
4171 | ac->ac_inode = ar->inode; | |
53accfa9 | 4172 | ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical; |
c9de560d AT |
4173 | ac->ac_o_ex.fe_group = group; |
4174 | ac->ac_o_ex.fe_start = block; | |
4175 | ac->ac_o_ex.fe_len = len; | |
53accfa9 | 4176 | ac->ac_g_ex = ac->ac_o_ex; |
c9de560d | 4177 | ac->ac_flags = ar->flags; |
c9de560d AT |
4178 | |
4179 | /* we have to define context: we'll we work with a file or | |
4180 | * locality group. this is a policy, actually */ | |
4181 | ext4_mb_group_or_file(ac); | |
4182 | ||
6ba495e9 | 4183 | mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, " |
c9de560d AT |
4184 | "left: %u/%u, right %u/%u to %swritable\n", |
4185 | (unsigned) ar->len, (unsigned) ar->logical, | |
4186 | (unsigned) ar->goal, ac->ac_flags, ac->ac_2order, | |
4187 | (unsigned) ar->lleft, (unsigned) ar->pleft, | |
4188 | (unsigned) ar->lright, (unsigned) ar->pright, | |
4189 | atomic_read(&ar->inode->i_writecount) ? "" : "non-"); | |
4190 | return 0; | |
4191 | ||
4192 | } | |
4193 | ||
6be2ded1 AK |
4194 | static noinline_for_stack void |
4195 | ext4_mb_discard_lg_preallocations(struct super_block *sb, | |
4196 | struct ext4_locality_group *lg, | |
4197 | int order, int total_entries) | |
4198 | { | |
4199 | ext4_group_t group = 0; | |
4200 | struct ext4_buddy e4b; | |
4201 | struct list_head discard_list; | |
4202 | struct ext4_prealloc_space *pa, *tmp; | |
6be2ded1 | 4203 | |
6ba495e9 | 4204 | mb_debug(1, "discard locality group preallocation\n"); |
6be2ded1 AK |
4205 | |
4206 | INIT_LIST_HEAD(&discard_list); | |
6be2ded1 AK |
4207 | |
4208 | spin_lock(&lg->lg_prealloc_lock); | |
4209 | list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order], | |
4210 | pa_inode_list) { | |
4211 | spin_lock(&pa->pa_lock); | |
4212 | if (atomic_read(&pa->pa_count)) { | |
4213 | /* | |
4214 | * This is the pa that we just used | |
4215 | * for block allocation. So don't | |
4216 | * free that | |
4217 | */ | |
4218 | spin_unlock(&pa->pa_lock); | |
4219 | continue; | |
4220 | } | |
4221 | if (pa->pa_deleted) { | |
4222 | spin_unlock(&pa->pa_lock); | |
4223 | continue; | |
4224 | } | |
4225 | /* only lg prealloc space */ | |
cc0fb9ad | 4226 | BUG_ON(pa->pa_type != MB_GROUP_PA); |
6be2ded1 AK |
4227 | |
4228 | /* seems this one can be freed ... */ | |
4229 | pa->pa_deleted = 1; | |
4230 | spin_unlock(&pa->pa_lock); | |
4231 | ||
4232 | list_del_rcu(&pa->pa_inode_list); | |
4233 | list_add(&pa->u.pa_tmp_list, &discard_list); | |
4234 | ||
4235 | total_entries--; | |
4236 | if (total_entries <= 5) { | |
4237 | /* | |
4238 | * we want to keep only 5 entries | |
4239 | * allowing it to grow to 8. This | |
4240 | * mak sure we don't call discard | |
4241 | * soon for this list. | |
4242 | */ | |
4243 | break; | |
4244 | } | |
4245 | } | |
4246 | spin_unlock(&lg->lg_prealloc_lock); | |
4247 | ||
4248 | list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) { | |
4249 | ||
bd86298e | 4250 | group = ext4_get_group_number(sb, pa->pa_pstart); |
6be2ded1 | 4251 | if (ext4_mb_load_buddy(sb, group, &e4b)) { |
12062ddd ES |
4252 | ext4_error(sb, "Error loading buddy information for %u", |
4253 | group); | |
6be2ded1 AK |
4254 | continue; |
4255 | } | |
4256 | ext4_lock_group(sb, group); | |
4257 | list_del(&pa->pa_group_list); | |
3e1e5f50 | 4258 | ext4_mb_release_group_pa(&e4b, pa); |
6be2ded1 AK |
4259 | ext4_unlock_group(sb, group); |
4260 | ||
e39e07fd | 4261 | ext4_mb_unload_buddy(&e4b); |
6be2ded1 AK |
4262 | list_del(&pa->u.pa_tmp_list); |
4263 | call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); | |
4264 | } | |
6be2ded1 AK |
4265 | } |
4266 | ||
4267 | /* | |
4268 | * We have incremented pa_count. So it cannot be freed at this | |
4269 | * point. Also we hold lg_mutex. So no parallel allocation is | |
4270 | * possible from this lg. That means pa_free cannot be updated. | |
4271 | * | |
4272 | * A parallel ext4_mb_discard_group_preallocations is possible. | |
4273 | * which can cause the lg_prealloc_list to be updated. | |
4274 | */ | |
4275 | ||
4276 | static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac) | |
4277 | { | |
4278 | int order, added = 0, lg_prealloc_count = 1; | |
4279 | struct super_block *sb = ac->ac_sb; | |
4280 | struct ext4_locality_group *lg = ac->ac_lg; | |
4281 | struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa; | |
4282 | ||
4283 | order = fls(pa->pa_free) - 1; | |
4284 | if (order > PREALLOC_TB_SIZE - 1) | |
4285 | /* The max size of hash table is PREALLOC_TB_SIZE */ | |
4286 | order = PREALLOC_TB_SIZE - 1; | |
4287 | /* Add the prealloc space to lg */ | |
f1167009 | 4288 | spin_lock(&lg->lg_prealloc_lock); |
6be2ded1 AK |
4289 | list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order], |
4290 | pa_inode_list) { | |
4291 | spin_lock(&tmp_pa->pa_lock); | |
4292 | if (tmp_pa->pa_deleted) { | |
e7c9e3e9 | 4293 | spin_unlock(&tmp_pa->pa_lock); |
6be2ded1 AK |
4294 | continue; |
4295 | } | |
4296 | if (!added && pa->pa_free < tmp_pa->pa_free) { | |
4297 | /* Add to the tail of the previous entry */ | |
4298 | list_add_tail_rcu(&pa->pa_inode_list, | |
4299 | &tmp_pa->pa_inode_list); | |
4300 | added = 1; | |
4301 | /* | |
4302 | * we want to count the total | |
4303 | * number of entries in the list | |
4304 | */ | |
4305 | } | |
4306 | spin_unlock(&tmp_pa->pa_lock); | |
4307 | lg_prealloc_count++; | |
4308 | } | |
4309 | if (!added) | |
4310 | list_add_tail_rcu(&pa->pa_inode_list, | |
4311 | &lg->lg_prealloc_list[order]); | |
f1167009 | 4312 | spin_unlock(&lg->lg_prealloc_lock); |
6be2ded1 AK |
4313 | |
4314 | /* Now trim the list to be not more than 8 elements */ | |
4315 | if (lg_prealloc_count > 8) { | |
4316 | ext4_mb_discard_lg_preallocations(sb, lg, | |
f1167009 | 4317 | order, lg_prealloc_count); |
6be2ded1 AK |
4318 | return; |
4319 | } | |
4320 | return ; | |
4321 | } | |
4322 | ||
c9de560d AT |
4323 | /* |
4324 | * release all resource we used in allocation | |
4325 | */ | |
4326 | static int ext4_mb_release_context(struct ext4_allocation_context *ac) | |
4327 | { | |
53accfa9 | 4328 | struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); |
6be2ded1 AK |
4329 | struct ext4_prealloc_space *pa = ac->ac_pa; |
4330 | if (pa) { | |
cc0fb9ad | 4331 | if (pa->pa_type == MB_GROUP_PA) { |
c9de560d | 4332 | /* see comment in ext4_mb_use_group_pa() */ |
6be2ded1 | 4333 | spin_lock(&pa->pa_lock); |
53accfa9 TT |
4334 | pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); |
4335 | pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); | |
6be2ded1 AK |
4336 | pa->pa_free -= ac->ac_b_ex.fe_len; |
4337 | pa->pa_len -= ac->ac_b_ex.fe_len; | |
4338 | spin_unlock(&pa->pa_lock); | |
c9de560d | 4339 | } |
c9de560d | 4340 | } |
ba443916 AK |
4341 | if (pa) { |
4342 | /* | |
4343 | * We want to add the pa to the right bucket. | |
4344 | * Remove it from the list and while adding | |
4345 | * make sure the list to which we are adding | |
44183d42 | 4346 | * doesn't grow big. |
ba443916 | 4347 | */ |
cc0fb9ad | 4348 | if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) { |
ba443916 AK |
4349 | spin_lock(pa->pa_obj_lock); |
4350 | list_del_rcu(&pa->pa_inode_list); | |
4351 | spin_unlock(pa->pa_obj_lock); | |
4352 | ext4_mb_add_n_trim(ac); | |
4353 | } | |
4354 | ext4_mb_put_pa(ac, ac->ac_sb, pa); | |
4355 | } | |
c9de560d AT |
4356 | if (ac->ac_bitmap_page) |
4357 | page_cache_release(ac->ac_bitmap_page); | |
4358 | if (ac->ac_buddy_page) | |
4359 | page_cache_release(ac->ac_buddy_page); | |
4360 | if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) | |
4361 | mutex_unlock(&ac->ac_lg->lg_mutex); | |
4362 | ext4_mb_collect_stats(ac); | |
4363 | return 0; | |
4364 | } | |
4365 | ||
4366 | static int ext4_mb_discard_preallocations(struct super_block *sb, int needed) | |
4367 | { | |
8df9675f | 4368 | ext4_group_t i, ngroups = ext4_get_groups_count(sb); |
c9de560d AT |
4369 | int ret; |
4370 | int freed = 0; | |
4371 | ||
9bffad1e | 4372 | trace_ext4_mb_discard_preallocations(sb, needed); |
8df9675f | 4373 | for (i = 0; i < ngroups && needed > 0; i++) { |
c9de560d AT |
4374 | ret = ext4_mb_discard_group_preallocations(sb, i, needed); |
4375 | freed += ret; | |
4376 | needed -= ret; | |
4377 | } | |
4378 | ||
4379 | return freed; | |
4380 | } | |
4381 | ||
4382 | /* | |
4383 | * Main entry point into mballoc to allocate blocks | |
4384 | * it tries to use preallocation first, then falls back | |
4385 | * to usual allocation | |
4386 | */ | |
4387 | ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle, | |
6c7a120a | 4388 | struct ext4_allocation_request *ar, int *errp) |
c9de560d | 4389 | { |
6bc6e63f | 4390 | int freed; |
256bdb49 | 4391 | struct ext4_allocation_context *ac = NULL; |
c9de560d AT |
4392 | struct ext4_sb_info *sbi; |
4393 | struct super_block *sb; | |
4394 | ext4_fsblk_t block = 0; | |
60e58e0f | 4395 | unsigned int inquota = 0; |
53accfa9 | 4396 | unsigned int reserv_clstrs = 0; |
c9de560d | 4397 | |
b10a44c3 | 4398 | might_sleep(); |
c9de560d AT |
4399 | sb = ar->inode->i_sb; |
4400 | sbi = EXT4_SB(sb); | |
4401 | ||
9bffad1e | 4402 | trace_ext4_request_blocks(ar); |
ba80b101 | 4403 | |
45dc63e7 DM |
4404 | /* Allow to use superuser reservation for quota file */ |
4405 | if (IS_NOQUOTA(ar->inode)) | |
4406 | ar->flags |= EXT4_MB_USE_ROOT_BLOCKS; | |
4407 | ||
60e58e0f MC |
4408 | /* |
4409 | * For delayed allocation, we could skip the ENOSPC and | |
4410 | * EDQUOT check, as blocks and quotas have been already | |
4411 | * reserved when data being copied into pagecache. | |
4412 | */ | |
f2321097 | 4413 | if (ext4_test_inode_state(ar->inode, EXT4_STATE_DELALLOC_RESERVED)) |
60e58e0f MC |
4414 | ar->flags |= EXT4_MB_DELALLOC_RESERVED; |
4415 | else { | |
4416 | /* Without delayed allocation we need to verify | |
4417 | * there is enough free blocks to do block allocation | |
4418 | * and verify allocation doesn't exceed the quota limits. | |
d2a17637 | 4419 | */ |
55f020db | 4420 | while (ar->len && |
e7d5f315 | 4421 | ext4_claim_free_clusters(sbi, ar->len, ar->flags)) { |
55f020db | 4422 | |
030ba6bc | 4423 | /* let others to free the space */ |
bb8b20ed | 4424 | cond_resched(); |
030ba6bc AK |
4425 | ar->len = ar->len >> 1; |
4426 | } | |
4427 | if (!ar->len) { | |
a30d542a AK |
4428 | *errp = -ENOSPC; |
4429 | return 0; | |
4430 | } | |
53accfa9 | 4431 | reserv_clstrs = ar->len; |
55f020db | 4432 | if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) { |
53accfa9 TT |
4433 | dquot_alloc_block_nofail(ar->inode, |
4434 | EXT4_C2B(sbi, ar->len)); | |
55f020db AH |
4435 | } else { |
4436 | while (ar->len && | |
53accfa9 TT |
4437 | dquot_alloc_block(ar->inode, |
4438 | EXT4_C2B(sbi, ar->len))) { | |
55f020db AH |
4439 | |
4440 | ar->flags |= EXT4_MB_HINT_NOPREALLOC; | |
4441 | ar->len--; | |
4442 | } | |
60e58e0f MC |
4443 | } |
4444 | inquota = ar->len; | |
4445 | if (ar->len == 0) { | |
4446 | *errp = -EDQUOT; | |
6c7a120a | 4447 | goto out; |
60e58e0f | 4448 | } |
07031431 | 4449 | } |
d2a17637 | 4450 | |
85556c9a | 4451 | ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS); |
833576b3 | 4452 | if (!ac) { |
363d4251 | 4453 | ar->len = 0; |
256bdb49 | 4454 | *errp = -ENOMEM; |
6c7a120a | 4455 | goto out; |
256bdb49 ES |
4456 | } |
4457 | ||
256bdb49 | 4458 | *errp = ext4_mb_initialize_context(ac, ar); |
c9de560d AT |
4459 | if (*errp) { |
4460 | ar->len = 0; | |
6c7a120a | 4461 | goto out; |
c9de560d AT |
4462 | } |
4463 | ||
256bdb49 ES |
4464 | ac->ac_op = EXT4_MB_HISTORY_PREALLOC; |
4465 | if (!ext4_mb_use_preallocated(ac)) { | |
256bdb49 ES |
4466 | ac->ac_op = EXT4_MB_HISTORY_ALLOC; |
4467 | ext4_mb_normalize_request(ac, ar); | |
c9de560d AT |
4468 | repeat: |
4469 | /* allocate space in core */ | |
6c7a120a | 4470 | *errp = ext4_mb_regular_allocator(ac); |
2c00ef3e AK |
4471 | if (*errp) |
4472 | goto discard_and_exit; | |
c9de560d AT |
4473 | |
4474 | /* as we've just preallocated more space than | |
2c00ef3e | 4475 | * user requested originally, we store allocated |
c9de560d | 4476 | * space in a special descriptor */ |
256bdb49 | 4477 | if (ac->ac_status == AC_STATUS_FOUND && |
2c00ef3e AK |
4478 | ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len) |
4479 | *errp = ext4_mb_new_preallocation(ac); | |
4480 | if (*errp) { | |
4481 | discard_and_exit: | |
4482 | ext4_discard_allocated_blocks(ac); | |
4483 | goto errout; | |
4484 | } | |
c9de560d | 4485 | } |
256bdb49 | 4486 | if (likely(ac->ac_status == AC_STATUS_FOUND)) { |
53accfa9 | 4487 | *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs); |
6c7a120a | 4488 | if (*errp == -EAGAIN) { |
8556e8f3 AK |
4489 | /* |
4490 | * drop the reference that we took | |
4491 | * in ext4_mb_use_best_found | |
4492 | */ | |
4493 | ext4_mb_release_context(ac); | |
519deca0 AK |
4494 | ac->ac_b_ex.fe_group = 0; |
4495 | ac->ac_b_ex.fe_start = 0; | |
4496 | ac->ac_b_ex.fe_len = 0; | |
4497 | ac->ac_status = AC_STATUS_CONTINUE; | |
4498 | goto repeat; | |
6d138ced | 4499 | } else if (*errp) { |
b844167e | 4500 | ext4_discard_allocated_blocks(ac); |
6d138ced ES |
4501 | goto errout; |
4502 | } else { | |
519deca0 AK |
4503 | block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); |
4504 | ar->len = ac->ac_b_ex.fe_len; | |
4505 | } | |
c9de560d | 4506 | } else { |
256bdb49 | 4507 | freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len); |
c9de560d AT |
4508 | if (freed) |
4509 | goto repeat; | |
4510 | *errp = -ENOSPC; | |
6c7a120a AK |
4511 | } |
4512 | ||
6d138ced | 4513 | errout: |
6c7a120a | 4514 | if (*errp) { |
256bdb49 | 4515 | ac->ac_b_ex.fe_len = 0; |
c9de560d | 4516 | ar->len = 0; |
256bdb49 | 4517 | ext4_mb_show_ac(ac); |
c9de560d | 4518 | } |
256bdb49 | 4519 | ext4_mb_release_context(ac); |
6c7a120a AK |
4520 | out: |
4521 | if (ac) | |
4522 | kmem_cache_free(ext4_ac_cachep, ac); | |
60e58e0f | 4523 | if (inquota && ar->len < inquota) |
53accfa9 | 4524 | dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len)); |
0087d9fb | 4525 | if (!ar->len) { |
f2321097 TT |
4526 | if (!ext4_test_inode_state(ar->inode, |
4527 | EXT4_STATE_DELALLOC_RESERVED)) | |
0087d9fb | 4528 | /* release all the reserved blocks if non delalloc */ |
57042651 | 4529 | percpu_counter_sub(&sbi->s_dirtyclusters_counter, |
53accfa9 | 4530 | reserv_clstrs); |
0087d9fb | 4531 | } |
c9de560d | 4532 | |
9bffad1e | 4533 | trace_ext4_allocate_blocks(ar, (unsigned long long)block); |
ba80b101 | 4534 | |
c9de560d AT |
4535 | return block; |
4536 | } | |
c9de560d | 4537 | |
c894058d AK |
4538 | /* |
4539 | * We can merge two free data extents only if the physical blocks | |
4540 | * are contiguous, AND the extents were freed by the same transaction, | |
4541 | * AND the blocks are associated with the same group. | |
4542 | */ | |
4543 | static int can_merge(struct ext4_free_data *entry1, | |
4544 | struct ext4_free_data *entry2) | |
4545 | { | |
18aadd47 BJ |
4546 | if ((entry1->efd_tid == entry2->efd_tid) && |
4547 | (entry1->efd_group == entry2->efd_group) && | |
4548 | ((entry1->efd_start_cluster + entry1->efd_count) == entry2->efd_start_cluster)) | |
c894058d AK |
4549 | return 1; |
4550 | return 0; | |
4551 | } | |
4552 | ||
4ddfef7b ES |
4553 | static noinline_for_stack int |
4554 | ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b, | |
7a2fcbf7 | 4555 | struct ext4_free_data *new_entry) |
c9de560d | 4556 | { |
e29136f8 | 4557 | ext4_group_t group = e4b->bd_group; |
84130193 | 4558 | ext4_grpblk_t cluster; |
7a2fcbf7 | 4559 | struct ext4_free_data *entry; |
c9de560d AT |
4560 | struct ext4_group_info *db = e4b->bd_info; |
4561 | struct super_block *sb = e4b->bd_sb; | |
4562 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
c894058d AK |
4563 | struct rb_node **n = &db->bb_free_root.rb_node, *node; |
4564 | struct rb_node *parent = NULL, *new_node; | |
4565 | ||
0390131b | 4566 | BUG_ON(!ext4_handle_valid(handle)); |
c9de560d AT |
4567 | BUG_ON(e4b->bd_bitmap_page == NULL); |
4568 | BUG_ON(e4b->bd_buddy_page == NULL); | |
4569 | ||
18aadd47 BJ |
4570 | new_node = &new_entry->efd_node; |
4571 | cluster = new_entry->efd_start_cluster; | |
c894058d | 4572 | |
c894058d AK |
4573 | if (!*n) { |
4574 | /* first free block exent. We need to | |
4575 | protect buddy cache from being freed, | |
4576 | * otherwise we'll refresh it from | |
4577 | * on-disk bitmap and lose not-yet-available | |
4578 | * blocks */ | |
4579 | page_cache_get(e4b->bd_buddy_page); | |
4580 | page_cache_get(e4b->bd_bitmap_page); | |
4581 | } | |
4582 | while (*n) { | |
4583 | parent = *n; | |
18aadd47 BJ |
4584 | entry = rb_entry(parent, struct ext4_free_data, efd_node); |
4585 | if (cluster < entry->efd_start_cluster) | |
c894058d | 4586 | n = &(*n)->rb_left; |
18aadd47 | 4587 | else if (cluster >= (entry->efd_start_cluster + entry->efd_count)) |
c894058d AK |
4588 | n = &(*n)->rb_right; |
4589 | else { | |
e29136f8 | 4590 | ext4_grp_locked_error(sb, group, 0, |
84130193 TT |
4591 | ext4_group_first_block_no(sb, group) + |
4592 | EXT4_C2B(sbi, cluster), | |
e29136f8 | 4593 | "Block already on to-be-freed list"); |
c894058d | 4594 | return 0; |
c9de560d | 4595 | } |
c894058d | 4596 | } |
c9de560d | 4597 | |
c894058d AK |
4598 | rb_link_node(new_node, parent, n); |
4599 | rb_insert_color(new_node, &db->bb_free_root); | |
4600 | ||
4601 | /* Now try to see the extent can be merged to left and right */ | |
4602 | node = rb_prev(new_node); | |
4603 | if (node) { | |
18aadd47 | 4604 | entry = rb_entry(node, struct ext4_free_data, efd_node); |
5d3ee208 DM |
4605 | if (can_merge(entry, new_entry) && |
4606 | ext4_journal_callback_try_del(handle, &entry->efd_jce)) { | |
18aadd47 BJ |
4607 | new_entry->efd_start_cluster = entry->efd_start_cluster; |
4608 | new_entry->efd_count += entry->efd_count; | |
c894058d | 4609 | rb_erase(node, &(db->bb_free_root)); |
18aadd47 | 4610 | kmem_cache_free(ext4_free_data_cachep, entry); |
c9de560d | 4611 | } |
c894058d | 4612 | } |
c9de560d | 4613 | |
c894058d AK |
4614 | node = rb_next(new_node); |
4615 | if (node) { | |
18aadd47 | 4616 | entry = rb_entry(node, struct ext4_free_data, efd_node); |
5d3ee208 DM |
4617 | if (can_merge(new_entry, entry) && |
4618 | ext4_journal_callback_try_del(handle, &entry->efd_jce)) { | |
18aadd47 | 4619 | new_entry->efd_count += entry->efd_count; |
c894058d | 4620 | rb_erase(node, &(db->bb_free_root)); |
18aadd47 | 4621 | kmem_cache_free(ext4_free_data_cachep, entry); |
c9de560d AT |
4622 | } |
4623 | } | |
3e624fc7 | 4624 | /* Add the extent to transaction's private list */ |
18aadd47 BJ |
4625 | ext4_journal_callback_add(handle, ext4_free_data_callback, |
4626 | &new_entry->efd_jce); | |
c9de560d AT |
4627 | return 0; |
4628 | } | |
4629 | ||
44338711 TT |
4630 | /** |
4631 | * ext4_free_blocks() -- Free given blocks and update quota | |
4632 | * @handle: handle for this transaction | |
4633 | * @inode: inode | |
4634 | * @block: start physical block to free | |
4635 | * @count: number of blocks to count | |
5def1360 | 4636 | * @flags: flags used by ext4_free_blocks |
c9de560d | 4637 | */ |
44338711 | 4638 | void ext4_free_blocks(handle_t *handle, struct inode *inode, |
e6362609 TT |
4639 | struct buffer_head *bh, ext4_fsblk_t block, |
4640 | unsigned long count, int flags) | |
c9de560d | 4641 | { |
26346ff6 | 4642 | struct buffer_head *bitmap_bh = NULL; |
c9de560d | 4643 | struct super_block *sb = inode->i_sb; |
c9de560d | 4644 | struct ext4_group_desc *gdp; |
498e5f24 | 4645 | unsigned int overflow; |
c9de560d AT |
4646 | ext4_grpblk_t bit; |
4647 | struct buffer_head *gd_bh; | |
4648 | ext4_group_t block_group; | |
4649 | struct ext4_sb_info *sbi; | |
4650 | struct ext4_buddy e4b; | |
84130193 | 4651 | unsigned int count_clusters; |
c9de560d AT |
4652 | int err = 0; |
4653 | int ret; | |
4654 | ||
b10a44c3 | 4655 | might_sleep(); |
e6362609 TT |
4656 | if (bh) { |
4657 | if (block) | |
4658 | BUG_ON(block != bh->b_blocknr); | |
4659 | else | |
4660 | block = bh->b_blocknr; | |
4661 | } | |
c9de560d | 4662 | |
c9de560d | 4663 | sbi = EXT4_SB(sb); |
1f2acb60 TT |
4664 | if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) && |
4665 | !ext4_data_block_valid(sbi, block, count)) { | |
12062ddd | 4666 | ext4_error(sb, "Freeing blocks not in datazone - " |
1f2acb60 | 4667 | "block = %llu, count = %lu", block, count); |
c9de560d AT |
4668 | goto error_return; |
4669 | } | |
4670 | ||
0610b6e9 | 4671 | ext4_debug("freeing block %llu\n", block); |
e6362609 TT |
4672 | trace_ext4_free_blocks(inode, block, count, flags); |
4673 | ||
4674 | if (flags & EXT4_FREE_BLOCKS_FORGET) { | |
4675 | struct buffer_head *tbh = bh; | |
4676 | int i; | |
4677 | ||
4678 | BUG_ON(bh && (count > 1)); | |
4679 | ||
4680 | for (i = 0; i < count; i++) { | |
2ed5724d | 4681 | cond_resched(); |
e6362609 TT |
4682 | if (!bh) |
4683 | tbh = sb_find_get_block(inode->i_sb, | |
4684 | block + i); | |
2ed5724d | 4685 | if (!tbh) |
87783690 | 4686 | continue; |
60e6679e | 4687 | ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA, |
e6362609 TT |
4688 | inode, tbh, block + i); |
4689 | } | |
4690 | } | |
4691 | ||
60e6679e | 4692 | /* |
e6362609 TT |
4693 | * We need to make sure we don't reuse the freed block until |
4694 | * after the transaction is committed, which we can do by | |
4695 | * treating the block as metadata, below. We make an | |
4696 | * exception if the inode is to be written in writeback mode | |
4697 | * since writeback mode has weak data consistency guarantees. | |
4698 | */ | |
4699 | if (!ext4_should_writeback_data(inode)) | |
4700 | flags |= EXT4_FREE_BLOCKS_METADATA; | |
c9de560d | 4701 | |
84130193 TT |
4702 | /* |
4703 | * If the extent to be freed does not begin on a cluster | |
4704 | * boundary, we need to deal with partial clusters at the | |
4705 | * beginning and end of the extent. Normally we will free | |
4706 | * blocks at the beginning or the end unless we are explicitly | |
4707 | * requested to avoid doing so. | |
4708 | */ | |
f5a44db5 | 4709 | overflow = EXT4_PBLK_COFF(sbi, block); |
84130193 TT |
4710 | if (overflow) { |
4711 | if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) { | |
4712 | overflow = sbi->s_cluster_ratio - overflow; | |
4713 | block += overflow; | |
4714 | if (count > overflow) | |
4715 | count -= overflow; | |
4716 | else | |
4717 | return; | |
4718 | } else { | |
4719 | block -= overflow; | |
4720 | count += overflow; | |
4721 | } | |
4722 | } | |
f5a44db5 | 4723 | overflow = EXT4_LBLK_COFF(sbi, count); |
84130193 TT |
4724 | if (overflow) { |
4725 | if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) { | |
4726 | if (count > overflow) | |
4727 | count -= overflow; | |
4728 | else | |
4729 | return; | |
4730 | } else | |
4731 | count += sbi->s_cluster_ratio - overflow; | |
4732 | } | |
4733 | ||
c9de560d AT |
4734 | do_more: |
4735 | overflow = 0; | |
4736 | ext4_get_group_no_and_offset(sb, block, &block_group, &bit); | |
4737 | ||
163a203d DW |
4738 | if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT( |
4739 | ext4_get_group_info(sb, block_group)))) | |
4740 | return; | |
4741 | ||
c9de560d AT |
4742 | /* |
4743 | * Check to see if we are freeing blocks across a group | |
4744 | * boundary. | |
4745 | */ | |
84130193 TT |
4746 | if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) { |
4747 | overflow = EXT4_C2B(sbi, bit) + count - | |
4748 | EXT4_BLOCKS_PER_GROUP(sb); | |
c9de560d AT |
4749 | count -= overflow; |
4750 | } | |
810da240 | 4751 | count_clusters = EXT4_NUM_B2C(sbi, count); |
574ca174 | 4752 | bitmap_bh = ext4_read_block_bitmap(sb, block_group); |
ce89f46c AK |
4753 | if (!bitmap_bh) { |
4754 | err = -EIO; | |
c9de560d | 4755 | goto error_return; |
ce89f46c | 4756 | } |
c9de560d | 4757 | gdp = ext4_get_group_desc(sb, block_group, &gd_bh); |
ce89f46c AK |
4758 | if (!gdp) { |
4759 | err = -EIO; | |
c9de560d | 4760 | goto error_return; |
ce89f46c | 4761 | } |
c9de560d AT |
4762 | |
4763 | if (in_range(ext4_block_bitmap(sb, gdp), block, count) || | |
4764 | in_range(ext4_inode_bitmap(sb, gdp), block, count) || | |
4765 | in_range(block, ext4_inode_table(sb, gdp), | |
84130193 | 4766 | EXT4_SB(sb)->s_itb_per_group) || |
c9de560d | 4767 | in_range(block + count - 1, ext4_inode_table(sb, gdp), |
84130193 | 4768 | EXT4_SB(sb)->s_itb_per_group)) { |
c9de560d | 4769 | |
12062ddd | 4770 | ext4_error(sb, "Freeing blocks in system zone - " |
0610b6e9 | 4771 | "Block = %llu, count = %lu", block, count); |
519deca0 AK |
4772 | /* err = 0. ext4_std_error should be a no op */ |
4773 | goto error_return; | |
c9de560d AT |
4774 | } |
4775 | ||
4776 | BUFFER_TRACE(bitmap_bh, "getting write access"); | |
4777 | err = ext4_journal_get_write_access(handle, bitmap_bh); | |
4778 | if (err) | |
4779 | goto error_return; | |
4780 | ||
4781 | /* | |
4782 | * We are about to modify some metadata. Call the journal APIs | |
4783 | * to unshare ->b_data if a currently-committing transaction is | |
4784 | * using it | |
4785 | */ | |
4786 | BUFFER_TRACE(gd_bh, "get_write_access"); | |
4787 | err = ext4_journal_get_write_access(handle, gd_bh); | |
4788 | if (err) | |
4789 | goto error_return; | |
c9de560d AT |
4790 | #ifdef AGGRESSIVE_CHECK |
4791 | { | |
4792 | int i; | |
84130193 | 4793 | for (i = 0; i < count_clusters; i++) |
c9de560d AT |
4794 | BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data)); |
4795 | } | |
4796 | #endif | |
84130193 | 4797 | trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters); |
c9de560d | 4798 | |
920313a7 AK |
4799 | err = ext4_mb_load_buddy(sb, block_group, &e4b); |
4800 | if (err) | |
4801 | goto error_return; | |
e6362609 TT |
4802 | |
4803 | if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) { | |
7a2fcbf7 AK |
4804 | struct ext4_free_data *new_entry; |
4805 | /* | |
4806 | * blocks being freed are metadata. these blocks shouldn't | |
4807 | * be used until this transaction is committed | |
4808 | */ | |
e7676a70 | 4809 | retry: |
18aadd47 | 4810 | new_entry = kmem_cache_alloc(ext4_free_data_cachep, GFP_NOFS); |
b72143ab | 4811 | if (!new_entry) { |
e7676a70 TT |
4812 | /* |
4813 | * We use a retry loop because | |
4814 | * ext4_free_blocks() is not allowed to fail. | |
4815 | */ | |
4816 | cond_resched(); | |
4817 | congestion_wait(BLK_RW_ASYNC, HZ/50); | |
4818 | goto retry; | |
b72143ab | 4819 | } |
18aadd47 BJ |
4820 | new_entry->efd_start_cluster = bit; |
4821 | new_entry->efd_group = block_group; | |
4822 | new_entry->efd_count = count_clusters; | |
4823 | new_entry->efd_tid = handle->h_transaction->t_tid; | |
955ce5f5 | 4824 | |
7a2fcbf7 | 4825 | ext4_lock_group(sb, block_group); |
84130193 | 4826 | mb_clear_bits(bitmap_bh->b_data, bit, count_clusters); |
7a2fcbf7 | 4827 | ext4_mb_free_metadata(handle, &e4b, new_entry); |
c9de560d | 4828 | } else { |
7a2fcbf7 AK |
4829 | /* need to update group_info->bb_free and bitmap |
4830 | * with group lock held. generate_buddy look at | |
4831 | * them with group lock_held | |
4832 | */ | |
d71c1ae2 LC |
4833 | if (test_opt(sb, DISCARD)) { |
4834 | err = ext4_issue_discard(sb, block_group, bit, count); | |
4835 | if (err && err != -EOPNOTSUPP) | |
4836 | ext4_msg(sb, KERN_WARNING, "discard request in" | |
4837 | " group:%d block:%d count:%lu failed" | |
4838 | " with %d", block_group, bit, count, | |
4839 | err); | |
8f9ff189 LC |
4840 | } else |
4841 | EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info); | |
d71c1ae2 | 4842 | |
955ce5f5 | 4843 | ext4_lock_group(sb, block_group); |
84130193 TT |
4844 | mb_clear_bits(bitmap_bh->b_data, bit, count_clusters); |
4845 | mb_free_blocks(inode, &e4b, bit, count_clusters); | |
c9de560d AT |
4846 | } |
4847 | ||
021b65bb TT |
4848 | ret = ext4_free_group_clusters(sb, gdp) + count_clusters; |
4849 | ext4_free_group_clusters_set(sb, gdp, ret); | |
79f1ba49 | 4850 | ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh); |
feb0ab32 | 4851 | ext4_group_desc_csum_set(sb, block_group, gdp); |
955ce5f5 | 4852 | ext4_unlock_group(sb, block_group); |
c9de560d | 4853 | |
772cb7c8 JS |
4854 | if (sbi->s_log_groups_per_flex) { |
4855 | ext4_group_t flex_group = ext4_flex_group(sbi, block_group); | |
90ba983f TT |
4856 | atomic64_add(count_clusters, |
4857 | &sbi->s_flex_groups[flex_group].free_clusters); | |
772cb7c8 JS |
4858 | } |
4859 | ||
71d4f7d0 | 4860 | if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE)) |
7b415bf6 | 4861 | dquot_free_block(inode, EXT4_C2B(sbi, count_clusters)); |
7d734532 JK |
4862 | percpu_counter_add(&sbi->s_freeclusters_counter, count_clusters); |
4863 | ||
4864 | ext4_mb_unload_buddy(&e4b); | |
7b415bf6 | 4865 | |
7a2fcbf7 AK |
4866 | /* We dirtied the bitmap block */ |
4867 | BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); | |
4868 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); | |
4869 | ||
c9de560d AT |
4870 | /* And the group descriptor block */ |
4871 | BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); | |
0390131b | 4872 | ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); |
c9de560d AT |
4873 | if (!err) |
4874 | err = ret; | |
4875 | ||
4876 | if (overflow && !err) { | |
4877 | block += count; | |
4878 | count = overflow; | |
4879 | put_bh(bitmap_bh); | |
4880 | goto do_more; | |
4881 | } | |
c9de560d AT |
4882 | error_return: |
4883 | brelse(bitmap_bh); | |
4884 | ext4_std_error(sb, err); | |
4885 | return; | |
4886 | } | |
7360d173 | 4887 | |
2846e820 | 4888 | /** |
0529155e | 4889 | * ext4_group_add_blocks() -- Add given blocks to an existing group |
2846e820 AG |
4890 | * @handle: handle to this transaction |
4891 | * @sb: super block | |
4907cb7b | 4892 | * @block: start physical block to add to the block group |
2846e820 AG |
4893 | * @count: number of blocks to free |
4894 | * | |
e73a347b | 4895 | * This marks the blocks as free in the bitmap and buddy. |
2846e820 | 4896 | */ |
cc7365df | 4897 | int ext4_group_add_blocks(handle_t *handle, struct super_block *sb, |
2846e820 AG |
4898 | ext4_fsblk_t block, unsigned long count) |
4899 | { | |
4900 | struct buffer_head *bitmap_bh = NULL; | |
4901 | struct buffer_head *gd_bh; | |
4902 | ext4_group_t block_group; | |
4903 | ext4_grpblk_t bit; | |
4904 | unsigned int i; | |
4905 | struct ext4_group_desc *desc; | |
4906 | struct ext4_sb_info *sbi = EXT4_SB(sb); | |
e73a347b | 4907 | struct ext4_buddy e4b; |
2846e820 AG |
4908 | int err = 0, ret, blk_free_count; |
4909 | ext4_grpblk_t blocks_freed; | |
2846e820 AG |
4910 | |
4911 | ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1); | |
4912 | ||
4740b830 YY |
4913 | if (count == 0) |
4914 | return 0; | |
4915 | ||
2846e820 | 4916 | ext4_get_group_no_and_offset(sb, block, &block_group, &bit); |
2846e820 AG |
4917 | /* |
4918 | * Check to see if we are freeing blocks across a group | |
4919 | * boundary. | |
4920 | */ | |
cc7365df YY |
4921 | if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) { |
4922 | ext4_warning(sb, "too much blocks added to group %u\n", | |
4923 | block_group); | |
4924 | err = -EINVAL; | |
2846e820 | 4925 | goto error_return; |
cc7365df | 4926 | } |
2cd05cc3 | 4927 | |
2846e820 | 4928 | bitmap_bh = ext4_read_block_bitmap(sb, block_group); |
cc7365df YY |
4929 | if (!bitmap_bh) { |
4930 | err = -EIO; | |
2846e820 | 4931 | goto error_return; |
cc7365df YY |
4932 | } |
4933 | ||
2846e820 | 4934 | desc = ext4_get_group_desc(sb, block_group, &gd_bh); |
cc7365df YY |
4935 | if (!desc) { |
4936 | err = -EIO; | |
2846e820 | 4937 | goto error_return; |
cc7365df | 4938 | } |
2846e820 AG |
4939 | |
4940 | if (in_range(ext4_block_bitmap(sb, desc), block, count) || | |
4941 | in_range(ext4_inode_bitmap(sb, desc), block, count) || | |
4942 | in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) || | |
4943 | in_range(block + count - 1, ext4_inode_table(sb, desc), | |
4944 | sbi->s_itb_per_group)) { | |
4945 | ext4_error(sb, "Adding blocks in system zones - " | |
4946 | "Block = %llu, count = %lu", | |
4947 | block, count); | |
cc7365df | 4948 | err = -EINVAL; |
2846e820 AG |
4949 | goto error_return; |
4950 | } | |
4951 | ||
2cd05cc3 TT |
4952 | BUFFER_TRACE(bitmap_bh, "getting write access"); |
4953 | err = ext4_journal_get_write_access(handle, bitmap_bh); | |
2846e820 AG |
4954 | if (err) |
4955 | goto error_return; | |
4956 | ||
4957 | /* | |
4958 | * We are about to modify some metadata. Call the journal APIs | |
4959 | * to unshare ->b_data if a currently-committing transaction is | |
4960 | * using it | |
4961 | */ | |
4962 | BUFFER_TRACE(gd_bh, "get_write_access"); | |
4963 | err = ext4_journal_get_write_access(handle, gd_bh); | |
4964 | if (err) | |
4965 | goto error_return; | |
e73a347b | 4966 | |
2846e820 AG |
4967 | for (i = 0, blocks_freed = 0; i < count; i++) { |
4968 | BUFFER_TRACE(bitmap_bh, "clear bit"); | |
e73a347b | 4969 | if (!mb_test_bit(bit + i, bitmap_bh->b_data)) { |
2846e820 AG |
4970 | ext4_error(sb, "bit already cleared for block %llu", |
4971 | (ext4_fsblk_t)(block + i)); | |
4972 | BUFFER_TRACE(bitmap_bh, "bit already cleared"); | |
4973 | } else { | |
4974 | blocks_freed++; | |
4975 | } | |
4976 | } | |
e73a347b AG |
4977 | |
4978 | err = ext4_mb_load_buddy(sb, block_group, &e4b); | |
4979 | if (err) | |
4980 | goto error_return; | |
4981 | ||
4982 | /* | |
4983 | * need to update group_info->bb_free and bitmap | |
4984 | * with group lock held. generate_buddy look at | |
4985 | * them with group lock_held | |
4986 | */ | |
2846e820 | 4987 | ext4_lock_group(sb, block_group); |
e73a347b AG |
4988 | mb_clear_bits(bitmap_bh->b_data, bit, count); |
4989 | mb_free_blocks(NULL, &e4b, bit, count); | |
021b65bb TT |
4990 | blk_free_count = blocks_freed + ext4_free_group_clusters(sb, desc); |
4991 | ext4_free_group_clusters_set(sb, desc, blk_free_count); | |
79f1ba49 | 4992 | ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh); |
feb0ab32 | 4993 | ext4_group_desc_csum_set(sb, block_group, desc); |
2846e820 | 4994 | ext4_unlock_group(sb, block_group); |
57042651 | 4995 | percpu_counter_add(&sbi->s_freeclusters_counter, |
810da240 | 4996 | EXT4_NUM_B2C(sbi, blocks_freed)); |
2846e820 AG |
4997 | |
4998 | if (sbi->s_log_groups_per_flex) { | |
4999 | ext4_group_t flex_group = ext4_flex_group(sbi, block_group); | |
90ba983f TT |
5000 | atomic64_add(EXT4_NUM_B2C(sbi, blocks_freed), |
5001 | &sbi->s_flex_groups[flex_group].free_clusters); | |
2846e820 | 5002 | } |
e73a347b AG |
5003 | |
5004 | ext4_mb_unload_buddy(&e4b); | |
2846e820 AG |
5005 | |
5006 | /* We dirtied the bitmap block */ | |
5007 | BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); | |
5008 | err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); | |
5009 | ||
5010 | /* And the group descriptor block */ | |
5011 | BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); | |
5012 | ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); | |
5013 | if (!err) | |
5014 | err = ret; | |
5015 | ||
5016 | error_return: | |
5017 | brelse(bitmap_bh); | |
5018 | ext4_std_error(sb, err); | |
cc7365df | 5019 | return err; |
2846e820 AG |
5020 | } |
5021 | ||
7360d173 LC |
5022 | /** |
5023 | * ext4_trim_extent -- function to TRIM one single free extent in the group | |
5024 | * @sb: super block for the file system | |
5025 | * @start: starting block of the free extent in the alloc. group | |
5026 | * @count: number of blocks to TRIM | |
5027 | * @group: alloc. group we are working with | |
5028 | * @e4b: ext4 buddy for the group | |
5029 | * | |
5030 | * Trim "count" blocks starting at "start" in the "group". To assure that no | |
5031 | * one will allocate those blocks, mark it as used in buddy bitmap. This must | |
5032 | * be called with under the group lock. | |
5033 | */ | |
d71c1ae2 | 5034 | static int ext4_trim_extent(struct super_block *sb, int start, int count, |
d9f34504 | 5035 | ext4_group_t group, struct ext4_buddy *e4b) |
e2cbd587 | 5036 | __releases(bitlock) |
5037 | __acquires(bitlock) | |
7360d173 LC |
5038 | { |
5039 | struct ext4_free_extent ex; | |
d71c1ae2 | 5040 | int ret = 0; |
7360d173 | 5041 | |
b3d4c2b1 TM |
5042 | trace_ext4_trim_extent(sb, group, start, count); |
5043 | ||
7360d173 LC |
5044 | assert_spin_locked(ext4_group_lock_ptr(sb, group)); |
5045 | ||
5046 | ex.fe_start = start; | |
5047 | ex.fe_group = group; | |
5048 | ex.fe_len = count; | |
5049 | ||
5050 | /* | |
5051 | * Mark blocks used, so no one can reuse them while | |
5052 | * being trimmed. | |
5053 | */ | |
5054 | mb_mark_used(e4b, &ex); | |
5055 | ext4_unlock_group(sb, group); | |
d71c1ae2 | 5056 | ret = ext4_issue_discard(sb, group, start, count); |
7360d173 LC |
5057 | ext4_lock_group(sb, group); |
5058 | mb_free_blocks(NULL, e4b, start, ex.fe_len); | |
d71c1ae2 | 5059 | return ret; |
7360d173 LC |
5060 | } |
5061 | ||
5062 | /** | |
5063 | * ext4_trim_all_free -- function to trim all free space in alloc. group | |
5064 | * @sb: super block for file system | |
22612283 | 5065 | * @group: group to be trimmed |
7360d173 LC |
5066 | * @start: first group block to examine |
5067 | * @max: last group block to examine | |
5068 | * @minblocks: minimum extent block count | |
5069 | * | |
5070 | * ext4_trim_all_free walks through group's buddy bitmap searching for free | |
5071 | * extents. When the free block is found, ext4_trim_extent is called to TRIM | |
5072 | * the extent. | |
5073 | * | |
5074 | * | |
5075 | * ext4_trim_all_free walks through group's block bitmap searching for free | |
5076 | * extents. When the free extent is found, mark it as used in group buddy | |
5077 | * bitmap. Then issue a TRIM command on this extent and free the extent in | |
5078 | * the group buddy bitmap. This is done until whole group is scanned. | |
5079 | */ | |
0b75a840 | 5080 | static ext4_grpblk_t |
78944086 LC |
5081 | ext4_trim_all_free(struct super_block *sb, ext4_group_t group, |
5082 | ext4_grpblk_t start, ext4_grpblk_t max, | |
5083 | ext4_grpblk_t minblocks) | |
7360d173 LC |
5084 | { |
5085 | void *bitmap; | |
169ddc3e | 5086 | ext4_grpblk_t next, count = 0, free_count = 0; |
78944086 | 5087 | struct ext4_buddy e4b; |
d71c1ae2 | 5088 | int ret = 0; |
7360d173 | 5089 | |
b3d4c2b1 TM |
5090 | trace_ext4_trim_all_free(sb, group, start, max); |
5091 | ||
78944086 LC |
5092 | ret = ext4_mb_load_buddy(sb, group, &e4b); |
5093 | if (ret) { | |
5094 | ext4_error(sb, "Error in loading buddy " | |
5095 | "information for %u", group); | |
5096 | return ret; | |
5097 | } | |
78944086 | 5098 | bitmap = e4b.bd_bitmap; |
28739eea LC |
5099 | |
5100 | ext4_lock_group(sb, group); | |
3d56b8d2 TM |
5101 | if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) && |
5102 | minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks)) | |
5103 | goto out; | |
5104 | ||
78944086 LC |
5105 | start = (e4b.bd_info->bb_first_free > start) ? |
5106 | e4b.bd_info->bb_first_free : start; | |
7360d173 | 5107 | |
913eed83 LC |
5108 | while (start <= max) { |
5109 | start = mb_find_next_zero_bit(bitmap, max + 1, start); | |
5110 | if (start > max) | |
7360d173 | 5111 | break; |
913eed83 | 5112 | next = mb_find_next_bit(bitmap, max + 1, start); |
7360d173 LC |
5113 | |
5114 | if ((next - start) >= minblocks) { | |
d71c1ae2 LC |
5115 | ret = ext4_trim_extent(sb, start, |
5116 | next - start, group, &e4b); | |
5117 | if (ret && ret != -EOPNOTSUPP) | |
5118 | break; | |
5119 | ret = 0; | |
7360d173 LC |
5120 | count += next - start; |
5121 | } | |
169ddc3e | 5122 | free_count += next - start; |
7360d173 LC |
5123 | start = next + 1; |
5124 | ||
5125 | if (fatal_signal_pending(current)) { | |
5126 | count = -ERESTARTSYS; | |
5127 | break; | |
5128 | } | |
5129 | ||
5130 | if (need_resched()) { | |
5131 | ext4_unlock_group(sb, group); | |
5132 | cond_resched(); | |
5133 | ext4_lock_group(sb, group); | |
5134 | } | |
5135 | ||
169ddc3e | 5136 | if ((e4b.bd_info->bb_free - free_count) < minblocks) |
7360d173 LC |
5137 | break; |
5138 | } | |
3d56b8d2 | 5139 | |
d71c1ae2 LC |
5140 | if (!ret) { |
5141 | ret = count; | |
3d56b8d2 | 5142 | EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info); |
d71c1ae2 | 5143 | } |
3d56b8d2 | 5144 | out: |
7360d173 | 5145 | ext4_unlock_group(sb, group); |
78944086 | 5146 | ext4_mb_unload_buddy(&e4b); |
7360d173 LC |
5147 | |
5148 | ext4_debug("trimmed %d blocks in the group %d\n", | |
5149 | count, group); | |
5150 | ||
d71c1ae2 | 5151 | return ret; |
7360d173 LC |
5152 | } |
5153 | ||
5154 | /** | |
5155 | * ext4_trim_fs() -- trim ioctl handle function | |
5156 | * @sb: superblock for filesystem | |
5157 | * @range: fstrim_range structure | |
5158 | * | |
5159 | * start: First Byte to trim | |
5160 | * len: number of Bytes to trim from start | |
5161 | * minlen: minimum extent length in Bytes | |
5162 | * ext4_trim_fs goes through all allocation groups containing Bytes from | |
5163 | * start to start+len. For each such a group ext4_trim_all_free function | |
5164 | * is invoked to trim all free space. | |
5165 | */ | |
5166 | int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range) | |
5167 | { | |
78944086 | 5168 | struct ext4_group_info *grp; |
913eed83 | 5169 | ext4_group_t group, first_group, last_group; |
7137d7a4 | 5170 | ext4_grpblk_t cnt = 0, first_cluster, last_cluster; |
913eed83 | 5171 | uint64_t start, end, minlen, trimmed = 0; |
0f0a25bf JK |
5172 | ext4_fsblk_t first_data_blk = |
5173 | le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block); | |
913eed83 | 5174 | ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es); |
7360d173 LC |
5175 | int ret = 0; |
5176 | ||
5177 | start = range->start >> sb->s_blocksize_bits; | |
913eed83 | 5178 | end = start + (range->len >> sb->s_blocksize_bits) - 1; |
aaf7d73e LC |
5179 | minlen = EXT4_NUM_B2C(EXT4_SB(sb), |
5180 | range->minlen >> sb->s_blocksize_bits); | |
7360d173 | 5181 | |
5de35e8d LC |
5182 | if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) || |
5183 | start >= max_blks || | |
5184 | range->len < sb->s_blocksize) | |
7360d173 | 5185 | return -EINVAL; |
913eed83 LC |
5186 | if (end >= max_blks) |
5187 | end = max_blks - 1; | |
5188 | if (end <= first_data_blk) | |
22f10457 | 5189 | goto out; |
913eed83 | 5190 | if (start < first_data_blk) |
0f0a25bf | 5191 | start = first_data_blk; |
7360d173 | 5192 | |
913eed83 | 5193 | /* Determine first and last group to examine based on start and end */ |
7360d173 | 5194 | ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start, |
7137d7a4 | 5195 | &first_group, &first_cluster); |
913eed83 | 5196 | ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end, |
7137d7a4 | 5197 | &last_group, &last_cluster); |
7360d173 | 5198 | |
913eed83 LC |
5199 | /* end now represents the last cluster to discard in this group */ |
5200 | end = EXT4_CLUSTERS_PER_GROUP(sb) - 1; | |
7360d173 LC |
5201 | |
5202 | for (group = first_group; group <= last_group; group++) { | |
78944086 LC |
5203 | grp = ext4_get_group_info(sb, group); |
5204 | /* We only do this if the grp has never been initialized */ | |
5205 | if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { | |
5206 | ret = ext4_mb_init_group(sb, group); | |
5207 | if (ret) | |
5208 | break; | |
7360d173 LC |
5209 | } |
5210 | ||
0ba08517 | 5211 | /* |
913eed83 LC |
5212 | * For all the groups except the last one, last cluster will |
5213 | * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to | |
5214 | * change it for the last group, note that last_cluster is | |
5215 | * already computed earlier by ext4_get_group_no_and_offset() | |
0ba08517 | 5216 | */ |
913eed83 LC |
5217 | if (group == last_group) |
5218 | end = last_cluster; | |
7360d173 | 5219 | |
78944086 | 5220 | if (grp->bb_free >= minlen) { |
7137d7a4 | 5221 | cnt = ext4_trim_all_free(sb, group, first_cluster, |
913eed83 | 5222 | end, minlen); |
7360d173 LC |
5223 | if (cnt < 0) { |
5224 | ret = cnt; | |
7360d173 LC |
5225 | break; |
5226 | } | |
21e7fd22 | 5227 | trimmed += cnt; |
7360d173 | 5228 | } |
913eed83 LC |
5229 | |
5230 | /* | |
5231 | * For every group except the first one, we are sure | |
5232 | * that the first cluster to discard will be cluster #0. | |
5233 | */ | |
7137d7a4 | 5234 | first_cluster = 0; |
7360d173 | 5235 | } |
7360d173 | 5236 | |
3d56b8d2 TM |
5237 | if (!ret) |
5238 | atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen); | |
5239 | ||
22f10457 | 5240 | out: |
aaf7d73e | 5241 | range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits; |
7360d173 LC |
5242 | return ret; |
5243 | } |